YOLOv8-CDD: A Salient Target Detection Model for Underwater Cultural Heritage in Complex Environments

Abstract

Existing object detection methods often experience a decline in accuracy when directly applied to underwater scenarios, making it difficult to meet the precision and reliability requirements of practical cultural heritage detection tasks. To address common challenges in underwater heritage detection-such as low image contrast, severe target occlusion, and insufficient detection accuracy-this paper proposes a novel detection model, YOLOv8-CDD. The model is designed to improve the recognition of salient target regions in complex underwater conditions, thereby enhancing the automation and efficiency of archaeological detection. YOLOv8-CDD is developed by deeply optimizing the backbone and detection head of the original YOLOv8 architecture. It integrates a C2f-DCNv4 structure, the CBAM attention mechanism, and the Dynamic4 detection head to significantly enhance feature extraction and the perception of small or occluded objects, while maintaining high computational efficiency. The model is systematically evaluated on both the public DUO dataset and a self-constructed underwater archaeological image dataset. Experimental results show that YOLOv8-CDD achieves an mAP@0.5 of 84.9% and an mAP@0.5:0.95 of 63.8% on the DUO dataset, representing improvements of 4.0% and 3.4%, respectively, over the baseline. On the self-constructed dataset, it attains an mAP@0.5 of 70.4%, with a gain of 4.1%.