TIDE-YOLO: Lightweight Algorithm for Underwater Object Detection

This study proposes TIDE-YOLO ( T RACON +  I nner-WIoU + Bi- D irectional FPN +  E MPC-Detect), a lightweight algorithm designed for underwater object detection (UOD) based on the YOLOv8s framework. This algorithm addresses several challenges commonly found in underwater environments, such as blurred images, the abundance of small objects with minimal distinguishing features, and the high computational requirements of models. Firstly, the T riple Attention Mechanism (TAM) and R eceptive-Field A ttention CON volution (RFAConv) are integrated into the C2f_bottleneck to design an enhanced C2f module called TRACON. This modification enhances the receptive field of the convolutional layer, thus improving the feature extraction of the model and its ability to detect small targets. Secondly, the Bi-directional Feature Pyramid Network (BiFPN) is used to enhance the model’s contextual information capture while reducing the parameter count. Thirdly, a lightweight and efficient detection head named EMPC-Detect, which integrates EMSConv and PConv, is proposed. EMPC-Detect improves the capability of the model to capture minute object details, all the while reducing both the parameter count and computational demands of the model. Finally, Inner-WIoU loss was designed by incorporating Inner-IoU and WIoU. Inner-WIoU replaced the CIoU loss to further improve the model’s accuracy and enhanced the algorithm's ability to generalize. TIDE-YOLO was assessed using DUO, UTDAC2020, and RUOD datasets, achieving an mAP50 scores of 87.1%, 86.0%, and 86.1%, respectively. Compared to YOLOv8s, TIDE-YOLO showed a substantial decrease in model size, parameter count, and computational demands, with a reduction of 67.9%, 73.4%, and 39.3%, respectively.