Underwater object detection algorithm integrating image enhancement and deformable convolution

Underwater biological detection plays a crucial role in the conservation of biodiversity and the exploration of underwater mineral resources. However, traditional object detection algorithms often suffer considerable performance degradation when confronted with underwater-specific challenges, including image blurring, small-scale targets, aggregation-induced occlusion, and irregular object shapes. To address these limitations, this study proposes a novel underwater object detection algorithm, DeformableConvModule-You Only Look Once (DCM-YOLO), based on YOLOv8s. First, to alleviate the degradation in underwater image quality, we incorporate an integrated image enhancement module, that is, UnitModule, which generates optimal input images for the detector. Second, the backbone network is redesigned via Deformable Convolution v4 (DCNv4), with targeted optimizations applied to enhance the model's capacity for detecting irregularly shaped objects and small-scale targets. In addition, the separated and enhancement attention module (SEAM) is integrated to better capture features of occluded targets. Finally, a dedicated detection head is added to further improve the model's ability to detect small objects. Comparative experiments and ablation experiments conducted on the Detecting Underwater Objects dataset validate the effectiveness of the proposed method. Specifically, the model improves the value of average precision (AP) from 60.8 % (for the YOLOv8s baseline model) to 65.0 %, and the AP at an intersection-over-union threshold of 0.5 ( AP 50 ) from 80.1 % (for the YOLOv8s baseline model) to 83.5 %. In addition, experimental evaluations on the URPC2020 and MS COCO2017 datasets confirm the advantages of the proposed DCM-YOLO algorithm. • The UnitModule is jointly trained with the detector, overcoming the decoupling between image enhancement and object detection, and guiding image enhancement toward directions favorable for detection. • DCNv4-based DCM blocks are designed to improve the backbone network's ability to extract features from irregularly shaped underwater organisms. • DySample and a specialized small object detection head are employed to enhance the network's sensitivity to small targets. • SEAM modules are introduced before each detection head to strengthen the model's responsiveness in densely occluded regions.