A lightweight underwater object detection with enhanced detail and edge-aware feature fusion

Underwater object detection often encounters challenges such as variable target scale, complex backgrounds, blurred object edges, and image distortion. In response to these challenges, a lightweight detection algorithm, EDFF-YOLO (Edge Detail Feature Fusion YOLO), is designed to enhance detection performance under these adverse conditions. To enhance the capability of the network to extract global features from images, a multi-scale residual enhancement module has been developed. This module captures and fuses a broader range of multi-scale contextual information. Secondly, a hybrid feature fusion module is proposed, which enhances the effectiveness of feature representation by using the hybrid local channel attention mechanism and element-wise operations to guide and fuse features. Then, a lightweight edge extraction block is designed to extract both edge and spatial information of the image, enriching feature diversity. Finally, the shared detail enhancement detection head is used to improve the ability of the detection head to capture details and to reduce the number of parameters and computational load of the algorithm. The experimental results reveal that the proposed algorithm outperforms the YOLOv8s baseline algorithm on the RUOD dataset. It demonstrates a reduction of 18 % and 30 % in both the number of parameters and computational load, respectively. Additionally, the [email protected] increases by 0.4 % to reach 88.1 %, surpassing the performance of other tested algorithms.