Underwater small target detection based on dynamic convolution and attention mechanism

Abstract

In ocean observation missions, unmanned autonomous ocean observation platforms play a crucial role, with precise target detection technology serving as a key support for the autonomous operation of unmanned platforms. Among various underwater sensing devices, side-scan sonar (SSS) has become a primary tool for wide-area underwater detection due to its extensive detection range. However, current research on target detection with SSS primarily focuses on large targets such as sunken ships and aircraft, lacking investigations into small targets. In this study, we collected data on underwater small targets using an unmanned boat equipped with SSS and proposed an enhancement method based on the YOLOv7 model for detecting small targets in SSS images. First, to obtain more accurate initial anchor boxes, we replaced the original k-means algorithm with the k-means++ algorithm. Next, we replaced ordinary convolution blocks in the backbone network with Omni-dimensional Dynamic Convolution (ODConv) to enhance the feature extraction capability for small targets. Subsequently, we inserted a Global Attention Mechanism (GAM) into the neck network to focus on global information and extract target features, effectively addressing the issue of sparse target features in SSS images. Finally, we mitigated the harmful gradients produced by low-quality annotated data by adopting Wise-IoU (WIoU) to improve the detection accuracy of small targets in SSS images. Through validation on the test set, the proposed method showed a significant improvement compared to the original YOLOv7, with increases of 5.05% and 2.51% in mAP@0.5 and mAP@0.5: 0.95 indicators, respectively. The proposed method demonstrated excellent performance in detecting small targets in SSS images and can be applied to the detection of underwater mines and small equipment, providing effective support for underwater small target detection tasks.

A Pruning and Distillation Based Compression Method for Sonar Image Detection Models

“…In the field of object detection, deep learning has achieved tremendous success due to its powerful feature extraction capabilities, effectively driving the development of this domain. Currently, convolutional neural network based object detection models have been improving in performance as their network depth increases, as seen in RCNN [1,2], SSD [3,4], and YOLO [5,6]. However, as the network depth of these models continues to grow, the storage space required for the models and the computational resources needed for inference also increase, posing challenges for specific tasks that cannot utilize high-performance processors.…”