Revision as of 06:26, 10 December 2024

Abstract

In exploring the performance-enhancing effects of compression sportswear, this study investigates its impact on trunk muscle activation during core stability training, with a particular focus on the biomechanical implications of this relationship. This research posits the existence of an Equilibrium Point of Motion (EPM) and employs surface Electromyography (EMG) to measure muscular responses in the Erector Spinae (ES), External Oblique (EO), and Rectus Abdominis (RA) during exercises performed with and without compression wear. Results revealed that compression garments significantly augment root mean square (RMS) values for ES and EO muscles, suggesting differential muscle activation patterns. Further, the application of OpenSim modeling of full-body musculoskeletal mechanics facilitated a novel analysis of muscle forces and activations, enhancing the understanding of sportswear functionality in an ergonomic context. This approach uncovered significant correlations between actual and simulated muscle activations, particularly for the ES and EO during a gluteal bridge exercise. These insights contribute to the development of ergonomically-engineered sportswear aimed at optimizing athletic performance and provide a valuable reference for future sportswear design and research.

Keywords: Sportswear; Model Simulation; Surface-Electromyography; Core training; Tight shorts; Compression

1. Introduction

Tight-fitting sportswear has received much attention due to its effect on the body's performance, which is related to the mechanics of contraction of pressure and activity [1,2]. The core stability training lies at the heart of many aspects of fitness and professional sports, in addition, to an evolving emphasis on functional integration for dynamic athletic performance [3].

Core stability training is a mode of exercise focused on strengthening the trunk muscles, which include the abdomen, back, pelvis, and hips, and also is clouded by sportswear shorts [4, 5, 6,7]. Two main methods to monitor or evaluate performance are subjective assessment of core stability through observation and objective measurement of biofeedback, including electromyography, motion capture, and further calculation or simulation of torque, and strength [8, 9]. Many researchers have supported some positive effects of the sportswear pressure impact on the overall impact of sportswear on the sensorimotor system and may decrease muscle soreness and pain [10, 11]. There remains a lack of detailed exploration into specific areas. Whereas, the effect mechanisms and extent of tight-fitting shorts are unclear to the performance of body core stability, focusing on the abdomen and waist region [12, 13]. To further explore the association between sportswear and body performance, comparing different wear conditions in experimental data.

Previous studies encompass the impact of wear conditions on core dynamics, ranging from the influence of footwear on muscle response time to the differentiated effects of wearable products on specific muscle groups [14,15]. Lee and Do distinguish tight-fitting shorts from compression stockings [16]. Their evaluation of the functions of these products, with a specific focus on the lower back during core training. This research suggests that the unified recruitment of abdominal and back muscles, coupled with additional support to erector spinae (ES), rectus abdominis (RA), and external oblique (EO), can be achieved through these wearables [17]. Maximal voluntary isometric contraction and peak muscle power after motion were greater than without compression garments with a slight effect [2] It provides a perspective on wearable products' potential impact on specific muscle groups during core training [18,19]. Because compression garment materials and pressures may alter the response to movement, previous reviews recommended that all compression garment studies measure and report pressures at multiple sites and specify the materials used [10,20].

While studies have reported positive effects, such as reduced muscle soreness and pain, there is a relative paucity of research on how these garments specifically affect core stability, particularly in the lower back and abdominal region. To address this gap, our study concentrates on a largely unexplored aspect: the activation patterns of core muscles, with an emphasis on the lumbar region, during core stability exercises while wearing compression shorts. This focus provides a novel vantage point by isolating the lower back area, a crucial element in core training, yet often overlooked in compression garment studies.

Despite extensive research on the biomechanical impact of tight sportswear, previous studies have predominantly centered on the generalized effects of sportswear pressure on muscle response and sensorimotor function. Significant gaps remain in our understanding, particularly concerning the activation patterns of muscles in the lower back and waist area during core stability exercises. This paper proposes to fill these gaps by exploring the specific biomechanical and ergonomic impacts of lower back compression afforded by tight-fitting shorts. This represents an innovative direction in textile performance research, with the potential to deepen our grasp of how compression garments influence core muscle activation, with implications for both athletic performance optimization and the design of sportswear that supports the lower back.

This research is pioneering in that it not only documents the muscle activation patterns but also quantifies the degree of compression and its direct correlation with muscle activation in the lumbar region. This duality in focus – examining both the physical pressure exerted by the garment and the physiological response of the body – marks a significant departure from previous work, which often neglects the lower back coverage and its specific activation during core exercises.

In recent years, in the field of research on sports biomechanics and muscle structure simulation, various methods and kinetic simulation software applications have provided insights into understanding muscle activity, muscle geometry, and muscle performance when subjected to loads. Stokes and Gardner-Morse constructed a mathematical model incorporating the abdominal muscles and performing optimization calculations with Matlab software [21]. It revealed how abdominal muscle activity reduces the load on the spine by increasing intra-abdominal pressure. Beaucage and Gauvreau estimated the forces on the lower lumbar spine during symmetric and asymmetric weightlifting tasks using OpenSim software [22]. The study revealed the specific effects of various tasks on spinal forces involving the gluteus maximus, psoas maximus, and paravertebral muscle groups, which provided new perspectives on the assessment of exercise mechanics loading.

Branyan and Menguc synthesized a personalized anatomical curve muscle geometry model to simulate the mechanical behavior of muscles such as latissimus dorsi, extensor spinae, rectus abdominis, and obliquus abdominis [23]. The geometric model of curved muscles was also mapped by incorporating muscle geometry, moment arms, and fiber direction paths. Guo and Ren refined and optimized the microfilm element embedded in a flexible multibody dynamics model to enhance the understanding of the role of the core musculature in regulating the pressure distribution and stability of the spine [24]. The model included core muscle groups such as rectus abdominis, obliques, adductors, and multifidus. In summary, these studies, through the use of quantitative measurements and model simulation techniques, provide a simulation basis for a comprehensive understanding of the behavior of human muscles during exercise states and the interactions between muscles and bones, which in turn helps to facilitate the optimization of sports equipment design. Key research results and major findings of previous researchers provide valuable references in the field of muscle biomechanics modeling and simulation analysis. This study seeks to reveal how varying levels of tightness and the anatomical fit of the shorts could modulate this muscle activity, thereby contributing to a more targeted approach in the design of compression sportswear. The ultimate aim is to provide empirical data that could lead to enhanced ergonomic design of sportswear, precisely engineered to improve core training outcomes and reduce injury risk associated with lower back strain. This paper advocated that the relationship between the compression of sportswear and subject ergonomic aids and mechanisms should be further enhanced by achieving an equilibrium of muscle activation and pressure level.

2. Methodology

2.1 Subjects

The experiment was approved by the institution’s human research ethics committee (EST-2024-001). Written informed consent was obtained from fifteen healthy male participants (22.4±2.7 years old, 68.5±1.6 kg, 178±2.2 cm). The study involved subjects' surface electromyography (EMG) during various exercises to measure muscle activity. EMG data were used to infer characteristics of muscular strength. All participants were healthy, and they had agreed to and understood the content of the experiment.

2.2. Standardized Compression

The study meticulously detailed the protocol implemented to ensure that each subject experienced a similar level of compression from the tight-fit shorts. Different sizes of shorts were provided to the subjects according to their somatotypes to ensure consistent clothing pressure. The clothing pressure experienced by the subjects was measured to confirm uniformity and compared against the pressure of cotton shorts, which served as a no-pressure control.

2.3. Instrumentation

2.3.1. Subjective Clothing Pressure Comfort Evaluation

Subjects were instructed to wear standardized tight-fit shorts for consistency. Following a tactile test, subjects completed a questionnaire regarding the comfort of the fabric. This subjective evaluation data was compiled and analyzed to compare comfort performance across various fabric types. Feedback on perceived compression was recorded, providing qualitative data to complement the quantitative analysis. Data analysis included assessing averaged pressure values and inter-subject variability to confirm a standardized perception of tightness among all participants.

2.3.2. Electromyography

Surface electromyography (EMG) was used to capture muscle activity during different core training exercises. EMG data were collected using the Delsys Trigno Avanti sensor hardware and EMGworks software (version EW4, Inc, USA) at a high sampling rate of 1092 Hz. Subjects participated under two wear conditions: tight-fitting sports shorts and cotton shorts without pressure. This aimed to explore the impact of sportswear on muscle activation. Table 1 provides an overview of the placement of EMG sensors on various muscles.

Table 1 Placement of EMG Sensors on Muscles in the Human Body

2.3.3. Testing Sport Shorts

The tight-fit shorts were selected and standardized to ensure each subject experienced uniform compression according to their somatotypes. The shorts' compression attributes were indicated by their elastic modulus in both warp and weft directions. Comparative analyses were conducted against control garments (cotton shorts with no compression features). This approach ensured a comparative analysis to discern potential differences in muscle activation patterns based on the type of sportswear worn during core exercises. Table 2 illustrates the fabric content of pants provided to participants.

Table 2 Detail of Various Garments Provided to Participants

2.3.4. Motion Capture

Inverse kinematics values for actual motion capture inputs in OpenSim. Optical motion capture system that captures data from the subject during hip bridge movements. When performing hip bridge motion capture experiments, the marker points were determined based on anatomical knowledge and in conjunction with the bony prominence points of the human body (see Figure 1).

Figure 1 Schematic diagram of reflective dots attached to the human body

2.4. Experiment Protocol

The experimental protocol was meticulously designed to ensure consistency, accuracy, and reproducibility in investigating EMG responses during core training exercises. Participants were briefed on the study's objectives and provided informed consent. Before starting the exercises, participants underwent a standardized warm-up routine to minimize injury risk and ensure optimal muscle activation. This routine included light activity and specific dynamic stretches targeting core muscle groups.

Participants performed a series of core training exercises (Table 3), including the gluteal bridge and seated leg raise [25]. Each exercise was repeated for a predetermined number of sets and repetitions. To control for rhythm and cadence during exercise execution, participants were provided with auditory cues from electronic metronomes set to 60 bpm. This control minimized variability in muscle activation patterns, ensuring observed differences were due to sportswear type rather than movement cadence [26].

Table 3 Demonstration of each exercise and details in number of sets, repetitions, and rest intervals

For the gluteal bridge (GB):

- Initial position: Subject lay on their back with knees bent, supported by heels, elbows, and head.

- Process: Lift the waist, arch the body so that hips are raised, and abdomen levels with knees, then slowly lower to the initial posture.

For the seated leg raise (SLR):

- Initial position: Lean back to maintain balance and keep the lower back straight.

- Process: Contract abdominal muscles, lean back slightly to maintain balance, and bend knees to lift legs [27].

The comprehensive experiment protocol, including standardized warm-up, exercise demonstration, controlled repetitions, and rhythm management, contributed to the robustness and validity of the observed EMG responses to varying sportswear conditions.

Data were analyzed using EMGworks Analysis software. EMG signals from six muscles were calculated in absolute value and filtered using a Butterworth bandwidth between 4 and 400 Hz. RMS procedures were applied, and data were subsetted by time point. Motion EMG data were averaged using SPSS to compare different wear conditions. Amplifiers (Trigno Wireless System, Delsys Inc., Boston, USA) were used for all participants. Peak and mean muscle activity were averaged for the right and left muscles. Ten repetitions were analyzed and averaged to form an "ensemble average" curve, summarizing typical movement characteristics. RMS was used for the time normalization of repetition cycles.

Analysis was performed using repeated measures ANOVA in SPSS (version, IBM Institute Inc., NC) with two within-subject factors (wear status, and exercise motion) [2]. As Mauchly’s test of sphericity did not satisfy the assumptions, the results were based on multivariate ANOVA. The study aimed to compare the impact of tight-fitting sports shorts on athlete performance during core training exercises, extending findings on the impact of sportswear on biomechanical aspects [9].

Our findings support functional sports device design and contribute to rehabilitation research with smart garments [28]. The repeated ANOVA of six EMG zones identified distinct patterns in motion curves when participants wore tight-fitting shorts compared to loose-fitting cotton underwear.

3. Results

3.1 Impact of Compression Garments on RMS

The examination of various wear statuses during core training revealed substantial variability in RMS values compared to instances involving cotton shorts and tight shorts. As depicted in Table 4, there were significant differences in almost all measurements  (RMSRA, RMSEO, and RMSES (sig=0.000<0.05) for both the wearing state and the motor task. the state of dress and the motor task had a significant effect on the measurements, with the specific effect size being measured by the partial Eta square. It is important to note that the observed power in Table 5 (Wear status: 285.534, Exercise motion: 2968.996) is also high, indicating that different exercise motions could be influential in our findings.

Table 4 Multivariate Tests of Muscle RMS Value Between-subjects Effects and In-subject Comparison Testing

Table 5 The Result of Multivariate Testing

In exercise motion, we observed pronounced enhancements in the Erector Spinae (EO) muscles during the Straight Leg Raise (SLR) regimen across all groups and exercises. This phenomenon was particularly decreased when participants wore tight-fitting shorts, underscoring the potential impact of clothing on muscle engagement [41]. The muscles in the back group, specifically the erector spinae (ES), exhibited increased recruitment of muscle units during the hip bridge (HB) exercise. The primary functional contribution was attributed to ES, with relatively lesser involvement of the rectus abdominis (RA) and external oblique (EO).In the initial phase of the seated leg raise exercise, the activation of the erector spinae(EO) in the control group wearing cotton shorts and rectus abdominis (RA) in the group wearing shorts is modest. Notably, towards the middle and end of the exercise sequence, there was a significant surge in EO activation with cotton pants. A parallel trend was observed in RA during the latter part of the exercise. In comparison to muscles without tight-fitting shorts, RA's root mean square (RMS) value in wearing tight shorts demonstrated a more balanced and consistent level throughout the entire duration. This suggests that the utilization of tights may contribute to a more measured and evenly distributed strength training experience.

3.2 Validation and analysis of muscle activation simulation results

The measured muscle activation was calculated using electromyography (EMG) data. The muscle forces calculated from the simulation were synchronized to correspond with the muscle activation values obtained from the EMG experiments. Each muscle was compared between the simulation-calculated muscle activation and the measured EMG activation. Trends and values of the two sets of results were analyzed. If the simulated activation shows a consistent trend with the measured EMG data and is numerically similar or can be reasonably calibrated to match, the model is considered to be relatively accurate.

The OpenSim model calculates the muscle joint moments by the static optimal solution method based on the results derived from reverse engineering [42,43]. The accuracy of the model and the consistency of the experimental data were assessed by constructing a model simulation of the core muscles of the low back in the hip-bridge movement (see Figure. 2, Figure. 3) and validating it in comparison with the experimental data of electromyography (EMG). The model including rectus abdominis, external abdominal obliques, and erector spinae was successfully modeled by the OpenSim multibody dynamics platform. The hip bridge maneuver was evaluated as an example to compare the trend of EMG signal curves with the simulated muscle force curves to verify the reasonableness of the established exercise biomechanics model.

Figure 2 Simulation of the core muscles during hip-bridge movements (initial state)

Figure 3 Simulation of the core muscles during hip-bridge (process state)

When comparing the ratio of the root-mean-square (RMS) value of electromyographic values during an actual movement to the maximum voluntary contraction (MVC), the unit usually used is the percentage, which represents the intensity of electromyographic activity as a relative value of the MVC. RMS, on the other hand, is usually measured in microvolts (µV), which directly reflects the change in amplitude of the electrical signal. For the muscle force data output from the OpenSim motion model, the units are usually Newtons (N). Since the strength of the EMG signal is not directly equivalent to the force produced by the muscle, a transformation or normalization method is required to correlate these two different measurements when making comparisons. Since the focus of the study was on the relative changes in muscle activity, this chapter chose to directly compare the trends of each of the two parts of the data without converting the force into units of the electrical signal. Despite the non-uniformity of the measures, the degree to which the model agrees with reality can be assessed qualitatively by analyzing the activity trend similarities and correlations.

In this study, the OpenSim multibody dynamics model was validated by comparing it with electromyography (EMG) experimental data. By comparing the muscle activation data in the model with the EMG data collected in the experiments one by one, the results showed that the two showed a high degree of consistency and maintained relatively close numerical proximity in terms of activation (Figure. 4, Figure.5, Figure. 6).

Figure 4 Comparison of EMG and simulated values of the erector spinae muscle

Figure 5 Comparison of EMG measurements of the external abdominal oblique muscle with simulated values

Figure 6 Comparison of electromyographic eigenvalues of the rectus abdominis muscle with simulated values

The lumbar back muscle force in the simulation calculation model corresponded to the root mean square amplitude data of EMG activities collected by the actual EMG.

4. Discussion

To investigate the impact of tight-fitting sports shorts on athlete performance during core training exercises, focusing on the electromyographic (EMG) activity of the rectus abdominis (RA), external oblique (EO), and erector spinae (ES) muscles. We established a novel measurement approach and aimed to contribute to the understanding of sportswear's ergonomic influence on musculoskeletal function in core training. Our methodology involved comparing EMG signals across different muscle groups, particularly emphasizing the potential variations in muscle activation patterns associated with the use of tight-fitting sports shorts during core training exercises. During core stability training, the RA, EO, and ES muscles were strengthened to achieve body stability. The effect mechanisms of the abdomen and lower back region are likely related to tights wearing [29]. It is therefore speculated that body core training muscle activation from focus is hardly negligible at the same level [30].

The findings presented a nuanced perspective on muscle activation. While no significant differences were observed in testing times for EMG activity in the RA group, the ES muscles exhibited weaker RMS volume and a flatter curve. This suggests a potential influence of sportswear on the engagement of specific muscle groups, particularly the ES muscles. Additionally, the study revealed intriguing patterns during the seat leg raise exercise, indicating that the engagement of the external oblique (EO) muscles was influenced by the characteristics of the tight-fitting shorts. This observation adds complexity to our understanding of how sportswear may affect muscles. The result may guide the design and development of compression garments for optimal performance and benefits during core training. The implications of our findings extend beyond the immediate scope of this study. The identified patterns in muscle activation provide valuable insights for the design of sportswear that aims to optimize muscle engagement during core training. This has practical applications not only in sports performance but also in rehabilitation, aligning with the growing interest in smart garment applications for health-related interventions. Furthermore, the observed differences in muscle activation patterns suggest that athletes and individuals engaging in core training exercises may benefit from sportswear tailored to specific muscle groups, potentially enhancing exercise performance and minimizing the risk of injury.

4.1. Influence of Compression Garments on RMS

There has been a limited focus on examining the functional effects of athletic shorts specifically tailored for core sports [8]. The neuromuscular efficiency of trunk muscles increases with trunk flexion (Weston et al., 2018). Consequently, one of the novel aspects of this research is its selection of a specific sport to conduct an in-depth analysis of the electromyography (EMG) activation patterns associated with wearing compression shorts during athletic performance [32,16]. The previous work has proved that pattern muscle movement of retraining. Results showed evidence that core exercise. However, the existing literature lacks a comprehensive investigation into the functionality of compression shorts during core sports activities [33,34].

To bridge this research gap, this study emphasizes and compares the effects of wearing compression shorts and non-compression shorts, specifically examining their influence on muscle activation patterns (as measured by EMG) during sporting activities. By evaluating the outcomes of this study, it becomes possible to establish the link between the results obtained and those of previous researchers, both their similarities and discrepancies. A similar conclusion of the overall results of the meta-analyses suggests a slight likelihood of the benefit of compression garments on exercise recovery [10,20]. The results suggest that compression garments may be recommended after resistance training or eccentric exercise to aid recovery of maximal strength, power, and endurance performance. Garment fabric deformation and generated internal stresses on the body have demonstrated that clothing pressure has a greater effect on relieving muscle fatigue. Our results are in general agreement with the previous study on wearing the tights status of the muscles implying a decreased  EMG amplitude and the peak torque of muscle contractions decreased approximately, compared to the control group [35]. Our current findings expand prior work with trunk muscle of RA, ES, EO activation mode, and tight shorts in core training, which derived a decrease in muscle strength and a lower EMG activity in the control group [14].

4.2 Equilibrium Point of Motion (EPM)

The findings of this research demonstrate that wearing compression shorts during core sports activities elicits specific EMG activation patterns in the muscles involved. These observations are consistent with the existing research on compression garments and their influence on muscle activation. Nam measured that a garment wear pressure above 4 kPa is regarded as harmful to health [36], by considering the broader implications of sportswear on various aspects of athletic performance—such as fatigue mitigation. Former findings suggested that wearing compression short-tights with a pressure intensity of 15-20 mm Hg at the thigh can reduce the development of fatigue in thigh muscles during submaximal running exercise in healthy adult males [37,38].

RMS values do not always rise with linear increases in elastic modulus. There was a tendency for excessively high thresholds of compression to affect muscle activity, resulting in a decrease in EMG RMS. As the meridional elastic modulus increases, the EMG signal may rise slightly and then fall. A higher modulus of elasticity initially increases the level of muscle activation but then decreases this activation.  

The point of motor equilibrium of muscle activation concerning time and stress. This equilibrium point may be a point in time and a specific stress value at which muscle activation is optimal or most adaptive. Moreover, this study provides support for the idea that compression shorts can enhance athletic performance by reducing muscle fatigue and improving exercise efficiency. The unique contribution lies in both the selected sport and the focused investigation of the effects of compression shorts on muscle activation.

4.3  Correlation analysis of simulated muscle force and electromyographic data

In this study, the statistical method of correlation analysis was used for consistency analysis to assess the strength of correlation between simulated and experimental data by calculating the Pearson correlation coefficient to assess the consistency between simulated and experimental data. The correlation between the model and the experimental real data is high. As can be seen from Table 6 below, correlation analysis was utilized to go over the correlation between the EMG measured values and the model output muscle strength, using the Pearson correlation coefficient to indicate the strength of the correlation. The correlation coefficient value between the measured and simulated values of the erector spinae, and rectus abdominis muscles of the gluteal bridge is 0.747 and shows significance at the 0.01 level, thus indicating a significant positive correlation between the two. The value of the correlation coefficient for external abdominal obliques was 0.455 and showed significance at a 0.05 level.

Table 6 Correlation analysis of analog and electromyographic eigenvalues of muscles

* p<0.05 ** p<0.01

In the statistical comparison plots (Figure 5, Figure 6, Figure 7), it can be observed that the overall trend of the simulated and measured data possesses a clear consistency, although a certain degree of error occurs in individual data points. The average error estimate from the overall analysis is around 8%, which shows that although there are minor deviations in the model, they are not enough to affect the validity of the model. Therefore, we can consider that the established virtual model can better reproduce and verify the rule of change of electromyography and its dynamic characteristics in the experiment. This model not only verifies the authenticity of the muscle force and activity trend but also can be used for further simulation and theoretical research, which provides a reliable tool for deepening the understanding of the mechanism of low back movement. The possible reasons for the errors on individual data points are attributed to two aspects. One is that the model simplifies the biomechanical system to a certain extent and may not fully reflect all detailed features. Parameters in the model are usually set based on mean values, which may differ from actual individuals. The experimental data reflect the physiological characteristics of a limited number of individuals, whereas the model is based on general anatomical and biomechanical parameter settings, and there may be bias due to individual differences. Second, EMG data acquisition in experiments may be affected by equipment limitations, noise interference, and other factors, resulting in data errors. Specific movements involve complex muscle synergies, and it is difficult for the model to fully capture the synergistic dynamics.

Simulation models based on similar schemes have been developed previously to analyze muscle vitality ratios. A comparison of computational and experimental results with this literature revealed that the measured EMG data showed activity levels of the lumbar muscles that matched the data obtained in the text [39.40].

5. Conclusions

1. Performance Improvement through Compression Shorts: Our study underscores the potential performance enhancement achievable by wearing compression short tights during exercise, particularly in activities involving core training. The application of appropriate pressure intensity was found to influence muscle activation patterns positively. We introduced a novel measurement approach to contribute to a deeper understanding of the ergonomic impact of sportswear on musculoskeletal function during core training.

2. Muscle Activation Patterns and Core Stability: Our methodology involved a comprehensive comparison of EMG signals across different muscle groups, with a specific focus on the RA, EO, and ES muscles during core stability training. The observed strengthening of these muscles might indicate improved core stability, emphasizing the efficacy of high waist compression shorts in targeted trunk muscle engagement.

3. Advocation equilibrium of muscle activation and compression from sportswear

The relationship between the compression of sportswear and subject ergonomic aids and mechanisms should be further enhanced by achieving an equilibrium of muscle activation and pressure level. This study provides support for the idea that compression shorts can enhance athletic performance by reducing muscle fatigue and improving exercise efficiency.

In the context of relevant CNN research, Sakai et al. [4] employed graph CNN to anticipate the pharmacological behavior of chemical structures, utilizing the constructed model for virtual screening and the identification of a novel serotonin transporter inhibitor. The efficacy of this new compound is akin to that of in-vitro- marketed drugs and has demonstrated antidepressant effects in behavioral analyses. Rački et al. [5] utilized CNN to investigate surface defects in solid oral pharmaceutical dosage forms. The proposed structural framework is evaluated for performance, demonstrating cutting-edge capabilities. The model attains remarkable performance with just 3% of parameter computations, yielding an approximate eight fold enhancement in drug property identification efficiency. Yoon et al. [6] employed CNN and generative adversarial networks to synthesize and explore colonoscopy imagery through the development of a proficiently trained system using imbalanced polyp data. Generative adversarial networks were harnessed to synthesize high-resolution comprehensive endoscopic images. The findings reveal that the system augmented with synthetic image enhancement exhibits a 17.5% greater sensitivity in image recognition. Benradi et al. [7] introduced a hybrid approach for facial recognition, melding CNN with feature extraction techniques. Empirical outcomes validate the method's efficacy in facial recognition, significantly enhancing precision and recall. Ahmad et al. [8] proposed a methodology for recognizing human activities grounded in deep temporal learning. The amalgamation of CNN features with bidirectional gated recurrent units, along with the application offeatureselection strategies, enhances accuracy and recall rates. Experimental results underscore the method's substantial practicality and accuracy in human activity recognition.

Furthermore, certain scholars have conducted relevant studies into museum exhibit attributes. Ahmad et al. [9] investigated the requisites of museums as perceived by the public, deriving insights from museum scholars and experts to outline the trajectory for developing museum exhibitions in Malaysia aimed at facilitating public learning. The results highlight the unique role of museum exhibits in residents' lifelong learning journey. Ryabinin and Kolesnik [10] harnessed a scientific visualization system to explore cyber- physical museum displays rooted in a system-on-a-chip architecture with a customized user interface. This research introduced an intelligent scientific visualization module capable of interactive engagement with and display of museum exhibits. Shahrizoda [11] delved into architectural and artistic solutions for museum exhibitions, clarifying architectural matters based on existing scientific and historical documents, and assessing prevailing characteristics of architectural and artistic solutions aligned with museum objectives.

These pertinent studies offer valuable points of reference and thought-provoking insights for the current research. They exemplify the diverse applications of CNN across domains such as drug research, image synthesis and recognition, and facial identification. These investigations underscore CNN's robust modeling and predictive capabilities, indicating its potential significance within the realm of cultural exhibits. They present intriguing concepts for image synthesis and enhancement within the museum exhibit domain. Additionally, in the arena of museum product development, the exploration of the integration of deep learning with other technologies holds promise for elevating the precision and effectiveness of product design.

2.2 An overview of the attractiveness of museum historical and cultural exhibits for humanistic care

Developing cultural tolerance in patient audiences based on the appeal of museum historical and cultural exhibits has been shown to be effective. In one study, an art museum educator instructed learners to observe museum artworks using the Visual Thinking Strategies (VTS) method. Of 407 students, 211 responded to a post-session survey (52%). 80% of respondents agreed or strongly agreed that ‘observing the art of the event was an effective means of initiating discussion with the interprofessional team’. Qualitative analysis of learner gains identified themes of open-mindedness, listening to other views and perspectives, collaboration/teamwork, patient-centredness and bias awareness [12]. The review also showed that art visualisation training through a local art museum hosted by an arts and culture educator was better able to improve the teaching of: facilitated observation, diagnostic skills, empathy, team building, communication skills, resilience and cultural sensitivity for medical students [13]. The limitations of these studies are that they are single-institution reports, short-term, involve small numbers of students, and often lack control. An analysis of 31 schools in the medical humanities at US medical schools found that 52% of the institutions surveyed had centres/departments focused on the humanities [14]. 65% offered 10 or more medical humanities-assisted courses per year, but only 29% offered a full immersion experience in the medical humanities. The study also found that the number of students in the medical humanities was not sufficiently high to justify a full immersion experience. Thus museum-based training has proven to be effective in the training of medical students, yet the amount of research and implementation is still lacking and is still in the beginning stages of exploration.

The cultural communication and emotional healing functions of museums are highly effective in caregiving. A study of 20 patients with cognitive disabilities (PWCI) and caregivers participated in a focus group at the Andy Warhol Museum of Art in an art engagement programme that included guided tours and art projects [15]. Research has revealed that patient humour and laughter in museum interventions, (Further study may reveal roles of humor and laughter in adaptation to cognitive decline and holistic interventions for improved quality of life.) can improve patients' ability to adapt to cognitive decline and holistic interventions for improved quality of life. Another study examined the feasibility of art museum trips as an intervention for chronic pain patients [16]. The art museum offered 1-hour docent-led tours to 54 individuals with chronic pain. 57% of participants reported pain relief during the tour, with an average pain relief of 47%. Subjects reported less social disconnection and pain unpleasantness before and after the trip. Respondents spoke of the isolating effects of chronic pain and how negative experiences with the healthcare system often exacerbate this sense of isolation. Participants viewed the art museum trip as a positive and inclusive experience with potential lasting benefits. Art museum trips are feasible for people with chronic pain, and participants reported positive effects on perceived social disconnection and pain.

Research related to the attractiveness of historical and cultural exhibits in museums has shown different developments in many fields, but there are serious gaps in combining intelligent environments to mitigate the impact of nurse audiences. With the development of information technology, the development and improvement of CNN algorithms have enabled the practical implementation of facial expression recognition technology in real life.The utilisation of DTs has facilitated the digital modelling of cultural displays. By applying attractiveness analysis models and artificial intelligence algorithms centred on historical and cultural exhibits, the attractiveness of historical and cultural exhibits in museums targeting humanistic contexts has been analysed, and with this foundation more directions of exploration and research can be expanded. This endeavour aims to increase the attractiveness of cultural exhibits while expanding the museum's cross-disciplinary reach and visibility to other fields.

3. Model research on the attractiveness of historical and cultural exhibits in the National Museum based on CNN

3.1 Functional analysis of cultural products in the National Museum of China

Viewed from a cultural design standpoint, the National Museum of China encompasses a historical journey spanning 5,000 years of traditional Chinese culture. Consequently, cultural products should meticulously select representative elements from the wealth of cultural resources, aligning with their distinctive historical significance. Through these cultural products, one can propagate informational culture within the cultural heritage, uphold esteemed traditional practices, and fortify the educational role of cultural offerings [17-19]. Firstly, the design of museum exhibition booths can consider isolating certain popular exhibits to disperse visitor traffic effectively. Additionally, in booth design, efforts should be made to encourage visitor engagement in interactive activities, fostering experiences deeply rooted in cultural impact. Regularly rotating less popular exhibits can align with the concept of catering to visitors. Given the subjective and challenging nature of exhibit evaluation, different visitors have varying cultural content preferences, inevitably leading to differences. A broader audience can be attracted by contemplating and crafting cultural art works, cultural attributes, and innovative choices, thus cultivating a positive feedback loop. The cultural products and functional attributes of the museum are delineated in Figure 1.

Figure 1. Structure of the cultural product features of the National Museum

The architectural framework of China National Museum's cultural product characteristics is illustrated in Figure 1. This framework comprises the input layer, data input embedding layer, standardization and normalization layer, output layer, and accuracy loss layer. The input layer receives data from cultural exhibit images as input. The data input embedding layer pre-processes and extracts features from input images, transforming them into a format the model can handle. The standardization and normalization layer processes the data for standardization and normalization, ensuring the model's universality across different datasets. The output layer is the model's final layer, used to predict the features of cultural exhibits. The accuracy loss layer calculates the error between the model's predicted and actual results, updating model parameters through back propagation. This process aims to minimize error loss and enhance model accuracy. Through the collaborative operation of these layers, the essential structural features of cultural products can be effectively extracted and abstracted, achieving automatic recognition, classification, and recommendation capabilities for cultural exhibits.

3.2 Analysis of DTs applied to digital modeling of museum historical and cultural exhibits

DTs can provide more realistic simulations of exhibition venues and exhibits, enabling visitors to understand better the context and meaning of exhibits [20,21]. This research applies DTs to the digital modeling of museum historical and cultural exhibits, including key element entity modeling, DTs virtual modeling, and virtual-real mapping association modeling.

In solid modeling, the elements involved in the activities include exhibit characteristics, functions, performance, etc. Considering the above factors, this research uses a formal modeling language to model the key elements of the exhibit digitization process. The model definition is shown in Eq. (1):

In Eq. (1), ${\textstyle PS}$ refers to the physical space of key elements in the digitization process of exhibits. ${\textstyle PE}$ refers to the collection of characteristics of historical and cultural exhibits in physical space. ${\displaystyle \vartriangleleft }$ refers to the collection of functions of historical and cultural exhibits in physical space. ${\textstyle PW}$ refers to the collection of historical and cultural exhibits in physical space. ${\textstyle PP}$ refers to the natural connection between ${\textstyle PE,\,PP}$, and ${\textstyle PW}$, indicating the natural interaction between the three. ${\textstyle PE,\,PP}$, and ${\textstyle PW}$ are all dynamic collections. Collection elements and their status can be continuously updated with the dynamic operation of the digitization process of exhibits.

DTs volume modeling needs high modularity, good scalability, and dynamic adaptability, which can be completed in information space using the parametric modeling method. Virtual models of physical entities are established in Tecnomatrix, Demo3D, Visual Components, and other software [22,23]. In addition to describing the geometric information and topological relationship of the automated production line, the virtual model also contains the complete dynamic engineering information description of each physical object [24,25]. The multiple- dimensional attributes of the model are parametrically defined to realize the real-time mapping of the digital modeling process of exhibits. The specific definition of DTs volume modeling is shown in Eq. (2):

where, ${\textstyle CS}$ refers to the information space of the key elements of the digitalization process of exhibits. ${\textstyle DE}$ refers to collecting characteristics of historical and cultural exhibits in the information space. ${\textstyle DW}$ refers to the collection of functions of historical and cultural exhibits in the information space. ${\displaystyle \vartriangleleft }$ refers to collecting historical and cultural exhibits in the information space. ${\textstyle DP}$ refers to the natural connection between ${\textstyle DE,\,DP}$, and ${\textstyle DW}$, indicating the natural interaction between the three. ${\textstyle DE,\,DP}$, and ${\textstyle DW}$ are all dynamic collections. The collection elements and their states in the information space are updated synchronously with the dynamic operation of the digitalization process of the exhibits in the physical space.

Finally, the virtual-real mapping association is modeled. The virtual-real mapping relationship between the two is further established based on establishing the physical space entity model ${\textstyle PS}$ and the information space twin model ${\textstyle CS}$. The formal modeling language is used to model its virtual-real mapping relationship. The model definition is shown in Eq. (3):

where, ${\displaystyle {\overset {1:1}{\Leftrightarrow }}}$ refers to the real two-way mapping between the physical space entity model and the information space twin model. ${\displaystyle \vartriangleleft }$ refers to the natural connections between different models. Therefore, ${\textstyle PE,DE,PP,DP,PW}$, and ${\textstyle PW}$ should maintain asynchronous two-way real.

3.3 Design and research of facial emotion recognition strategy based on CNN

Emotion refers to the subjective emotional experience and internal psychological state that humans have in response to external stimuli. It encompasses various complex experiences and reactions such as happiness, sadness, surprise, anger, fear, and more. Facial expressions are the outward manifestation of emotions, reflecting the emotional states of individuals through muscle movements and neural transmission. Facial expressions can be automatically recognized and classified using computer vision and image processing techniques, helping people gain a more accurate understanding of visitors' emotional states and responses in front of historical and cultural exhibits in museums. Facial recognition technology can identify facial expressions automatically by analyzing the changes and combinations of facial features, such as the eyes, mouth, eyebrows, and more. For example, when visitors appreciate an exhibit in a museum, they may display different emotions and attitudes, such as liking, disliking, or surprise. By utilizing facial recognition technology, people can monitor visitors'emotional changes and responses in real-time and connect this information with the knowledge and stories related to the displayed historical and cultural artifacts. This enhances visitors' experience and cultural awareness. In conclusion, facial expressions are the outward manifestation of emotions, and facial recognition technology can infer visitors'emotional states and responses by analyzing these expressions. By employing these technologies, people can better understand visitors' emotional experiences and responses in front of historical and cultural exhibits in museums, providing a more scientific and effective reference for exhibition planning and cultural heritage preservation in museums.

Since the input data source of tourist facial expressions is image sequence information collected by surveillance cameras, the model must maintain high recognition accuracy while maintaining high recognition speed [26]. The mini_xception model boasts a lightweight design, enabling it to maintain relatively high accuracy while reducing the network's parameter count and computational load. This lightweight structure makes the mini_xception network well-suited for integration into practical applications, particularly in resource- constrained scenarios [27,28]. Therefore, mini_xception is chosen as the basic network structure. Firstly, layers are added in batches after each convolution to speed up training and improve the model's generalization ability. A regularized dropout rate is used in the model to randomly eliminate some nodes in the network with a given probability. While the network structure is simplified, over-adjustment of the network is avoided. The classification ability of the network is improved. The principle is that in forwarding propagation, some neurons'activation values are randomly stopped in a certain proportion. Therefore, the model is less dependent on some local features. The actual visitors ofthe museum are different, and the lighting intensity of the booth is also different. The main pre-processing methods adopted include grayscale, normalization, and histogram processing. Based on CNN, the structural framework of the designed facial emotion recognition model is shown in Figure 2.

Figure 2. Framework of facial emotion recognition model structure based on Mini_Xception and ResNet

Assuming there is a sequence offacial images of a group oftourists encompassing diverse emotional expressions as they stand in front of museum exhibits. The objective is to employ a model for recognizing and classifying these emotions while maintaining a balance between speed and accuracy. The mini_xception architecture has been chosen for the network structure. Following each convolutional layer, batch normalization layers have been added to expedite training and enhance model generalization. Additionally, dropout regularization has been employed within the model to deactivate some nodes, mitigating the risk of overfitting randomly. Given a sequence of facial images of different tourists, during the forward propagation phase, a portion of neuron activations is randomly halted to a certain proportion. This ensures the model does not excessively rely on specific localized features. Considering the varying illumination conditions faced by real tourists in the museum, pre-processing is essential for input facial images. Operations such as cropping, scaling, and grayscale conversion are conducted to optimize the image input for network processing. Subsequently, the mini_xception algorithm is applied to extract features from the facial images. This algorithm accelerates recognition speed while reducing the number of parameters and computations. The mini_xception algorithm transforms tourists' facial images into feature-rich representations encompassing spatial and semantic information. Following this, the ResNet algorithm is employed to fuse features extracted by mini_xception. ResNet, being aresidual network, leverages residual connections to reduce training time and enhance accuracy. This algorithm combines the spatial and semantic features derived from mini_xception to elevate emotion classification's accuracy and generalization capabilities. Ultimately, a Softmax classifier is utilized to categorize the fused features, thereby predicting the emotional category of the facial image. This classifier maps the feature vector into a probability space, yielding the predicted probabilities for each category. Suppose the model has been trained. In that case, a facial image of a tourist can be predicted as belonging to the "liking" emotional category. From capturing facial images to feature extraction and emotional classification, this algorithm aids in the automated recognition of tourists' emotional expressions before historical and cultural exhibits in museums. Consequently, it offers a more scientific and effective basis for museum exhibit planning and cultural heritage preservation.

In this model, ${\textstyle W}$ and ${\textstyle I}$ represent one network layer's weight tensor and input tensor, respectively. Among them, ${\textstyle c}$ represents the channel, ${\displaystyle w<w_{in}{\mbox{ }},h<h_{in}}$ represents the width and height, respectively. The network approximates ${\textstyle W}$ through a binary tensor ${\displaystyle B{\text{ e}}\{+1,-1\}^{c{\text{х}}w{\text{х}}h}}$ ${\displaystyle cxwxh}$ and a scaling factor ${\displaystyle C{\mbox{ }}eR+}$, as shown in Eq. (4):

where, ${\textstyle W}$ refers to the real weight.

In order to obtain the best approximation, it is assumed that vectors w and b represent weights and binarized weights, respectively, as shown in Eq. (5):

The determination of the optimal scale factor and binarization parameters is shown in Eq. (6):

The two sides of the above formula are subtracted, and the optimization function is obtained, as shown in Eqs. (7) and (8):

where, ${\displaystyle C^{\ast },b^{\ast }}$ represent the scale factor and the weight optimization value after binarization, respectively. Eq. (7) is expanded to obtain Eq. (9):

where, ${\displaystyle be\{+1,-1\}^{n}}$, ${\displaystyle b^{T}b=n}$ is a constant. ${\displaystyle w}$ refers to a known quantity, so ${\displaystyle w^{T}w}$ is also a constant. Let ${\displaystyle q=w^{T}w}$, Eq. (9) can be simplified to Eq. (10):

${\displaystyle C}$ refers to a positive value. The optimization problem of ${\displaystyle b}$ is transformed into Eq. (11):

According to the characteristics of the value of ${\displaystyle b}$, Eq. (12) is calculated:

In order to obtain the optimal value of ${\displaystyle C^{\ast }}$, ${\displaystyle J}$ takes the partial derivative of ${\displaystyle C}$ and makes it zero. The optimization result of ${\displaystyle C}$ is obtained, as shown in Eq. (13):

The value of ${\displaystyle b^{\ast }}$ is plugged into Eqs. (13), and Eq. (14) is obtained:

In order to further reduce the quantization error, the network's weight and activation quantization process is introduced into a scaling factor. Quantization weights ${\textstyle K_{b}(w)}$ and activations ${\textstyle K_{b}(C)}$, including scale factors, are obtained, as shown in Eqs. (15) and (16):

In Eq. (16), ${\textstyle b,h}$ are the results of the weight vector ${\textstyle w}$ and the activation vector ${\displaystyle C}$ after the Sign function. ${\displaystyle C,\beta }$ is the corresponding scaling factor. The definition of the convolution operation of the network forward process is shown in Eq. (17):

In Eq. (17), ${\textstyle z}$ refers to the output and ${\displaystyle \sigma \left(\cdot \right)}$ refers to the activation function. ${\displaystyle h{\text{e}}\left\{+1,-1\right\}^{n}}$ refers to the activation value after binarization. ${\displaystyle \beta }$ is the scaling factor for the activation.${\textstyle t}$ is the convolution operation. ⊙ points to the inner product of the quantity.

The expression of the approximate point product between ${\textstyle C}$ and ${\textstyle w}$ is shown in Eq. (18):

The optimization problem to be solved is shown in Eq. (19):

The weight vectors are optimized, and the emotions of faces exhibited by DTs are analyzed precisely.

In order to solve the optimization problem defined by Eq. (19) using swarm intelligence optimization algorithms, the steps are as follows:

Step 1: Randomly generate a set of particles, where each particle represents a possible solution in the solution space. Each particle has its own position and velocity, initialized with appropriate values.

Step 2: Calculate the fitness value of each particle based on the objective function defined in Eq. (19). The fitness value indicates the quality of the solution corresponding to the particle.

Step 3: For each particle, update its individual best position based on its current fitness value and historical best fitness value.

Step 4: Update the global best position of the swarm based on the fitness values of all particles. Record the optimal solution and its fitness value.

Step 5: Update the velocity and position of each particle based on its current velocity, individual best position, and global best position.

Step 6: Check if the termination condition is met, such as reaching the maximum number of iterations or the fitness value reaching a predetermined threshold. If the condition is met, proceed to step 8; otherwise, continue with steps 2 to 6.

Step 7: Repeat steps 3 to 6 until the termination condition is met.

Step 8: The optimal solution recorded during the iteration process is the optimization result of the problem.

3.4 Evaluation model of the attractiveness of cultural exhibits based on face recognition algorithm

The data recognized by the face recognition algorithm is mainly used to associate the subsequent facial expression information with character recognition. The residence time of each visitor in the current state can also be calculated through identity recognition information. Since the representation of features in consecutive images is difficult to change in the actual computational process, continuous recognition inevitably leads to many repeated expression results. The facial expression network information is extracted every 15 frames. Facial expression classification theory and created datasets are used as the basis. The client's facial expressions are divided into six categories. Simple results from facial expression recognition may not accurately reflect visitors' satisfaction with an exhibit. Therefore, information about the expressions of tourists is provided in the index. The attractiveness evaluation model of cultural exhibits is analyzed. The evaluation process of the attractiveness of cultural products is shown in Figure 3.

Figure 3. Evaluation process of the attraction of museum historical and cultural products

In Figure 3, the evaluation process of the attractiveness of historical and cultural artifacts in museums can be divided into five steps. Step 1: Model Initialization. An appropriate model needs to be selected, and its parameters and network structure are initialized. Commonly used models include CNN and Recurrent Neural Network (RNN). Step 2: Computing Hidden Layer and Unit Outputs. The selected model is used to compute the hidden layer and unit outputs by inputting the relevant data of historical and cultural artifacts in museums. Step 3: Data Input and Vectorization. The relevant data of historical and cultural artifacts in museums is transformed into vector form for computation and processing. Step 4: Calculating Deviation between Target and Prediction. A comparison is made between the predicted results and the actual results to calculate the deviation or error between them. Step 5: Output and Weight Updating. The model parameters are updated through backpropagation based on the magnitude of the error. This aims to minimize the error loss and improve the accuracy ofthe model. This process can be performed in both forward and backward directions, continuously optimizing the model's performance by updating weights and biases using the backpropagation algorithm. In summary, through the iterative process ofthese five steps, the accuracy and precision of the attractiveness evaluation model for historical and cultural artifacts in museums can be gradually improved. This provides a more scientific and effective basis for museum exhibition planning and cultural heritage preservation.

3.5 Data transmission and experimental research

The experiment is conducted in a museum located in City B. The duration of the experiment is one month. Firstly, a random sample of 500 visitors who visited the museum was selected as the subjects. Then, cameras are installed in various exhibition halls to collect facial expression data of the visitors using facial recognition technology in front of different exhibits. The data is acquired by installing cameras in the museum's ten booths. When the file is saved, the identification information ofthe exposure point in the filename is used to distinguish the video data of each exposure point. Ifthe traffic is high, it is limited to the amount of data to be considered, and 10 minutes of exposure input data plus traffic is selected for each program. The CNN algorithm detects faces in each frame of video data. Then, a 160*160-pixel face image is cut out, and the face edge is output at a specific position ofthe frame image. After the visitor obtains the collected face image, each point is input with an identification code. From the data in the MySQL database, quantitative scoring information, human identification information, residence time information, and exposure point information of the program representation labels are imported. Front- end code is used to connect to the back-end. Then, JavaScript is used to access the database data. Finally, an object library that displays attractiveness metrics is used to display the data.

In addition, in order to further analyze the attractiveness of the museum's historical and cultural exhibits, the proposed model algorithm is compared with the performance of Mini_Xception, ResNet, and Benradiet al. (2023). The recognition accuracy, scalability rate, data transmission delay of the system, and the recognition speed of tourists' emotions are analyzed using the human-computer interaction (HCI) system. Additionally, the higher the attraction ofthe exhibits, the longer the visitors stay in the exhibition area. The system analyzes the facial expression scores of tourists in front of different numbered exhibits and the visitor's stay time. It compares the attractiveness of different numbered exhibits according to the facial expression score and the visitor's stay time.

4. Results and discussion

4.1 Comparison of recognition accuracy and expansion rate of DTs system

In order to analyze the performance of each algorithm, different algorithms are used in the DTs system to analyze the accuracy of tourist emotion recognition and the system's scalability, as shown in Figure 4.

Figure 4. Recognition accuracy and scalability curves of different algorithms in museum exhibit DTs system (a. recognition accuracy curve; b.scalability rate change curve)

In Figure 4, as the model step size parameter increases, the accuracy of different algorithms in the DTs system for emotion recognition will slowly increase. When the model step size parameter is 0, the proposed algorithm's recognition accuracy is 75.60%, while the recognition accuracy of other models is less than 73.67%. When the model step size parameter is increased to 100, the recognition accuracy of the proposed model algorithm reaches 95.92%, while other model algorithms do not exceed 90.39%. The recognition accuracy of each model algorithm is sorted from large to small: proposed model algorithm> Benradi et al. (2023)> Mini_Xception> ResNet. In addition, when the model step size parameter is 80, the proposed algorithm improves to 95.75%, which is better than other model algorithms. The results of comparative experiments show that the proposed recognition method can greatly improve the recognition accuracy and extension rate. The recognition accuracy can be increased by 5.53% compared with existing methods. When visitors identify DTs exhibits,the model can achieve more accurate emotion recognition.

4.2 Comparison of results between data transmission delay and tourist emotion recognition speed

The system uses different algorithms to compare the data transmission delay time and the speed of tourist emotion recognition, as shown in Figure 5.

Figure 5. Data transmission delay time and tourist emotion recognition speed change curve of different algorithms in the system (a) Data transmission delay time. (b) Tourist emotion recognition speed change curve)

In Figure 5, in the system, as the number ofmodel iterations increases, the data transmission delay time of each model algorithm shows a downward trend. When the model iterates 100 times, the data transmission delay of the traditional algorithm ResNet is 3.96s. The data transmission delay of the improved CNN model is only 2.71s. After 550 model iterations, the data transmission delay of the traditional algorithm ResNet drops to 2.49s. The data transmission delay of the optimized CNN algorithm can be reduced to 1.67s. In addition, the emotion recognition speed of the traditional algorithm ResNet is 3.22s after 550 iterations, and the emotion recognition speed ofthe improved CNN algorithm is only 1.33s. The data analysis results show that the data transmission delay ofthe proposed model algorithm can be reduced to 2.71s. The algorithm is superior to the traditional algorithm regarding emotion recognition speed.

In addition, the resource consumption of the four models is compared, as shown in Figure 6.

Figure 6. Comparison of Computational Resource Consumption for Different Algorithms

In Figure 6, the proposed algorithm's computational time, memory usage, and GPU memory consumption are 4.2s, 120MB, and 500MB, respectively, the lowest values among the four algorithms. This data indicates that the proposed algorithm has advantages over other models in terms of computational time, memory usage, and GPU memory consumption. The result means that the proposed algorithm can efficiently utilize computational resources and has lower resource consumption in the task of recognizing the emotions of museum visitors.

Furthermore, the research also compared the results of different algorithms in terms ofdata transmission latency and the speed ofvisitor emotion recognition. The findings revealed that with an increase in the number of model iterations, the data transmission latency for each algorithm exhibited a decreasing trend. For instance, after 100 rounds of model iteration, the data transmission latency for the conventional ResNet algorithm was 3.96 seconds, whereas the improved CNN model demonstrated a data transmission latency of only 2.71 seconds. Upon further augmentation to 550 rounds of model iteration, the data transmission latency of the conventional ResNet algorithm decreased to 2.49 seconds, while the optimized CNN algorithm achieved a reduced data transmission latency of 1.67 seconds. Similarly, the emotion recognition speed of the conventional ResNet algorithm was 3.22 seconds after 550 iterations, whereas the improved CNN algorithm exhibited a recognition speed ofmerely 1.33 seconds. Based on the results of data analysis, the proposed model algorithm is capable ofreducing data transmission latency to 2.71 seconds and outperforms the traditional algorithm in terms of emotion recognition speed.

4.3 Analysis of facial expression scores and visitor stay time results for exhibits with different numbers

The facial expression scores of tourists in front of different numbered exhibits and the data on tourists' staying time in the exhibition area are analyzed, as shown in Figure 7.

Figure 7. Average score of facial expressions of museum visitors to different exhibits and the average stay time change

In Figure 7, the horizontal axis represents the museum's numbered historical and cultural exhibits. The left vertical axis corresponds to the average facial expression scores of 500 visitors in front of various exhibits, while the right vertical axis represents the visitors' average duration ofstay and admiration time in front ofthese exhibits. There is a certain variation in both the average facial expression scores and the average duration of stay for visitors in front of different museum exhibits. Exhibit 4 displays a higher average facial expression score (73) and a relatively longer average stay time (106 seconds). In contrast, exhibit 5 exhibits a lower average facial expression score (46) along with a relatively shorter average stay time (39 seconds). This observation implies a potential correlation between visitors' facial expressions and their duration of stay when encountering different exhibits.

Within this experiment, emotional reactions are the feelings experienced by visitors while observing the exhibits. These emotional reactions may encompass excitement, pleasure, curiosity, and surprise. Such emotional responses could directly impact visitors' preferences for the exhibits. For instance, if a visitor experiences strong excitement and pleasure when observing a particular exhibit, they are likely to favor it. Facial expressions serve as physiological manifestations of emotional experiences conveyed through facial muscle movements like smiling or frowning. Different facial expressions could be associated with distinct emotions and moods, which, in turn, may affect visitors'preferences for the exhibits. For instance, a visitor displaying a noticeable smile while observing an exhibit may indicate their pleasure and satisfaction with it. Positive emotional experiences associated with an exhibit may lead to a higher degree of liking. Comparing visitors'facial expressions in front of different exhibits can help determine which exhibits align better with their interests and preferences. Moreover, visitors might exhibit positive facial expressions such as smiling or widening eyes when they come across exhibits they like. These positive expressions could signify their interest or satisfaction with the exhibit, prompting them to spend more time better appreciating it. Conversely, negative facial expressions like furrowing brows or frowns might indicate dissatisfaction, disinterest, or confusion, leading to reduced interest and shorter stays. Therefore, the experiment compares visitors' average facial expression scores and their average duration ofstay. The result shows a significance level (P) of 0.0138, which is less than the conventional threshold of 0.05, indicating a statistically significant positive correlation between visitors' facial expressions and their duration of stay. In this experiment, the research explored the correlation between facial expression scores and dwell time among visitors by recognizing and analyzing visitors' facial expressions. Facial expression scores were evaluated based on facial muscle movements to assess visitors' emotional experiences, while dwell time referred to the time visitors spent viewing exhibits. The research results demonstrated a statistically significant positive correlation between facial expression scores and dwell time. This implies that when visitors have higher facial expression scores, they tend to spend a longer time in front of exhibits. A higher facial expression score might indicate a positive emotional experience, such as excitement, curiosity, or satisfaction, prompting them to develop a greater fondness for the exhibit and be willing to invest more time in appreciation. Furthermore, the value of the correlation coefficient provides a deeper understanding. Based on the magnitude and sign of the correlation coefficient, the degree and direction ofthe association between facial expression scores and dwell time can be determined. As an illustrative instance ofthis research, the obtained results revealed a correlation coefficient of 0.0138, falling below the established conventional threshold of 0.05. The data indicate a significant positive correlation between the two, implying that as facial expression scores increase, visitors' dwell time also correspondingly increases.

Therefore, the findings of this research reveal a close relationship between visitors' facial expression scores and their dwell time within the exhibition hall. This discovery holds important implications for exhibition design and improvement, aiding in enhancing the attractiveness of exhibits and visitor experiences, thereby elevating visitor satisfaction and the overall success ofthe exhibition.

4.4 Discussion

In summary, the proposed model exhibits low latency in data transmission and improves the accuracy of emotion recognition. Furthermore, a comparison was made with other studies. Razzaq et al. [29] proposed DeepClass-Rooms, a DTs framework for attendance and course content monitoring in public sector schools in Punjab, Pakistan. It employed high-end computing devices with readers and fog layers for attendance monitoring and content matching. CNN was utilized for on-campus and online courses to enhance the educational level. Sun et al. [30] introduced a novel technique that formalizes personality as a DTs model by observing users' posting content and liking behavior. A multi-task learning deep neural network (DNN) model was employed to predict users' personalities based on two types of data representations. Experimental results demonstrated that combining these two types of data could improve the accuracy of personality prediction. Lin and Xiong [31] proposed a framework for performing controllable facial editing in video reconstruction. By retraining the generator of a generative adversarial network, a novel personalized generative adversarial network inversion was proposed for real face embeddings cropped from videos, preserving the identity details of real faces. The results indicate that this method achieves notable identity preservation and semantic disentanglement in controllable facial editing, surpassing recent state- of-the-art methods. In conclusion, an increasing body of research suggests that combining DTs technology with DNN can enhance the accuracy and efficiency of facial expression recognition.

5. Conclusion

With the rapid advancement of science and technology, AI and neural network algorithms have pervaded diverse domains across society. The integration of information technology has profoundly impacted daily life through the meticulous analysis of extensive datasets. Within the context of disseminating historical and cultural narratives, conventional museums often rely on staff members' prolonged visual observations and intuitive assessments to gauge the appeal of exhibits to visitors. This qualitative approach, while competent, poses challenges in terms of quantifiability. This research leverages DTs to digitally model historical and cultural artifacts digitally, enhancing the CNN algorithm to construct a facial emotion recognition model. By coupling this with visitors'dwell time within the exhibit areas, the degree of attraction exerted by each booth upon tourists is assessed and quantified. The functional analysis of historical and cultural artifacts at the museum provides insight. However, certain limitations persist. Foremost among these is the need for enhancement in the accuracy of the facial expression recognition algorithm. Subsequent research endeavors will seek to realize real-time detection of changes in visitors' facial expressions and dwell time through the refinement of the face detection algorithm, thereby addressing the appeal of cultural exhibits to tourists via intelligent HCI approaches.

The combination of museums and medical care is a relatively new and noteworthy direction, especially in terms of cultural sensitivity, mental health and emotional healing, and museums can have a unique effect. The current research is still in the stage of the preliminary attempt to combine history and culture, and the way and effect of the combination of museum and medical treatment are still being explored.

References