Improved YOLOv7 model for underwater sonar image object detection

Abstract

Underwater object detection technology for sonar images is widely employed and in high demand for civil and military purposes. However, due to the inhomogeneity of the seawater medium, which causes attenuation and distortion of the acoustic signal, achieving ideal performance for sonar object detection approaches is challenging. Furthermore, the process of acoustic wave transmission is further complicated by floating objects and particles, resulting in increased multipath effects. This study proposes an object detection method named YOLOv7C, which is based on deep learning to address these challenges. An attention mechanism is incorporated into the backbone network of the model to improve its ability to handle complex backgrounds in sonar images and effectively extract features. In addition, to facilitate high-order interaction between key features, the Neck part of the network integrates Multi-GnBlock blocks. Model redundancy pruning is used to substantially reduce the model size while maintaining high detection accuracy, thereby enhancing real-time performance. The proposed YOLOv7C achieves a 1.9% increase in the mean average precision, reduces the model memory by 47.50% after pruning, and enhances the detection speed by nearly 2.5 times compared with the YOLOv7C. These findings indicate that the success rate of multi-object detection is significantly enhanced by the attention mechanism and the new module. Additionally, the model can be controlled within a reasonable size by employing appropriate pruning methods.