A YOLO-Based Object Detection Method for Forward-Looking Sonar Images: A Case Study of YOLO-CSA

To address the challenges of low resolution, interference from complex underwater environments, and low accuracy in target detection for sonar images, this paper proposes an improved YOLOv5s-based detection model. First, the GSConv module replaces standard convolution in the neck network, reducing model complexity while improving recognition accuracy. Second, by integrating the SA attention mechanism, the proposed CSA module substitutes the original C3 module, enhancing focus on target features while maintaining lightweight and flexible design. Experiments were conducted on a self-built underwater sonar image dataset. Results demonstrate that the improved model achieves mAP50, mAP50-95, and Recall values of $83.1 \%, 36.4 \%$, and $76.8 \%$, respectively, representing improvements of $2.6 \%, 0.8 \%$, and $4.0 \%$ over the baseline YOLOv5s. Comparative experiments with other YOLO series algorithms (e.g., YOLOv8s, YOLOv11s) further confirm significant performance gains. Additionally, the proposed model reduces parameters by 0.4M, effectively balancing lightweight design and detection performance. This work provides a practical real-time solution for forward-looking sonar image target detection.