GCT-YOLOv5: a lightweight and efficient object detection model of real-time side-scan sonar image

Due to the high imaging resolution and large detection area, side-scan sonar (SSS) has a wide range of applications in underwater detection, such as shipwreck and wrecked aircraft search and rescue. However, limited by complex underwater environments and insufficient detection equipment performance, it is still very difficult to achieve real-time automatic object detection during the search and rescue process. To solve these problems, a new lightweight SSS image object detection model, GCT-YOLOv5, is proposed in this paper. Firstly, the C3Ghost and Ghost module are integrated into the proposed model to reduce the computational complexity. Subsequently, the coordinate attention mechanism is embedded within the backbone network to extract crucial features of objects and suppress the interference of background noise, thereby enhancing detection accuracy. Finally, transposed convolution is adopted in the neck network to achieve higher up-sampling performance and further enhance the ability of object feature perception. The experimental results demonstrate that compared to YOLOv5s, GCT-YOLOv5 has 48% fewer number of parameters and 45% fewer FLOPs. Furthermore, mAP@0.5 and mAP@0.5:0.95 of GCT-YOLOv5 are raised by 3.1 and 1.5%, respectively. GCT-YOLOv5 outperforms various commonly used object detection algorithms of SSS image in terms of accuracy and speed. In general, the GCT-YOLOv5 has the characteristics of strong robustness, lightweight, and high efficiency, which is especially suitable for situations requiring real-time detection based on SSS image. The source code and dataset of GCT-YOLOv5 can be viewed at https://github.com/gx-123/GCT-YOLOv5.