CSTC-YOLOv8: underwater object detection model based on improved YOLOv8 for side scan sonar images

Side scan sonar (SSS) object detection is critical for ocean exploration, yet existing models struggle with challenges, such as noise interference, blurred edges, and multiscale variability in SSS images, which hinder effective feature extraction and accurate detection. To address these limitations, we propose CSTC-YOLOv8, an advanced object detection model that integrates coordinate attention (CA), Swin Transformer (ST), and cross-stage partial spatial pyramid pooling group (CSPSPPG) mechanisms into the YOLOv8 architecture. CSTC-YOLOv8 introduces a coordinate attention module to enhance feature extraction by encoding spatial and channel information, improving the detection of small and complex objects. The integration of ST and CSPSPPG modules strengthens multiscale feature representation, enabling the model to handle significant scale variations in SSS imagery effectively. To ensure practical deployment in underwater systems, the model is further optimized through knowledge distillation, model compression, and TensorRT acceleration, resulting in a lightweight version suitable for real-time applications. The proposed framework was rigorously evaluated on a comprehensive SSS image dataset, achieving state-of-the-art performance with mAP50 and mAP50:95 scores of 82.6% and 47.4%, respectively. These results represent significant improvements of 15% and 8.7% over the baseline YOLOv8 model. Moreover, the optimized lightweight variant demonstrates exceptional computational efficiency, achieving an inference speed of 29.4 frames per second, which corresponds to a 38% enhancement in processing throughput. This performance advancement, coupled with the model’s improved detection accuracy, establishes CSTC-YOLOv8 as a robust solution for real-time underwater target detection applications.