A Novel Lightweight Model for Underwater Small Object Detection

Underwater object detection is widely used in underwater resource exploration and fisheries management. However, the complex underwater environment and limited computational resources pose significant challenges for accurate and efficient object detection. To address these challenges, a lightweight and efficient underwater object detection model, YOLO-SBL, is proposed in this paper. First, a novel SDGSConv convolution is introduced to mitigate the loss of fine-grained information caused by strided convolutions. Second, inspired by the Bidirectional Feature Pyramid Network (BiFPN) approach, the weighted fusion of features at different scales is achieved. Finally, a lightweight shared detail enhancement detection head (LSDED) is proposed, which enhances feature extraction capabilities while reducing the number of parameters. On the UPRC2020 dataset, YOLO-SBL achieves a [email protected] of 84.1%, improving by 1.4% over the original YOLOv8n model. The parameter count of the model is reduced by 20%, and the FPS increases from 240 to 249. On the NVIDIA Jetson TX2 edge device, the inference speed was reduced from 65 milliseconds to 51 milliseconds, meeting real-time requirements.