Abstract

Real-time object detection requires identifying objects in video streams or consecutive images with minimal latency, yet it continues to struggle with non-salient objects—those that are small, occluded, or otherwise inconspicuous. To address this limitation, this paper proposes CAM-YOLO, an enhanced architecture based on YOLOv8. First, to mitigate the baseline model’s limited representational capacity for non-salient targets, we introduce a Multi-Scale Aggregation Module (MSAM) into the feature fusion process, enabling the backbone network to extract more discriminative fine-grained features. Second, to better capture global contextual relationships associated with such objects, we design a Contextual Association Module (CAM) that explicitly models long-range spatial dependencies. Furthermore, we integrate a Dual-Branch Attention Mechanism (DBAM) to refine the feature processing flow, thereby strengthening the contextual feature representations crucial for detecting non-salient instances. Extensive experiments on two large-scale public benchmarks, Microsoft Common Objects in Context 2017 (MS COCO 2017) and PASCAL Visual Object Classes (PASCAL VOC), demonstrate that CAM-YOLO achieves highly competitive performance compared to several widely-adopted realtime detectors.OPEN ACCESS Received: 19/11/2025 Accepted: 15/01/2026 Published: 21/07/2026