SSE-YOLO : A Lighter and Faster Object Detection Network for Small Targets in Sonar Images

Due to the small size of the targets in sonar images, state-of-the-art target detection algorithms hardly make adaptable improvements to the sonar image. This paper proposes a lightweight target detection network, namely SSE-YOLO, to meet the premise of real-time fast detection of underwater targets under equipment computing power limitations. This improved model designs a lightweight feature extraction module SSE based on YOLOV3 to build a new lightweight backbone network. In order to enhance small target characterization, a new YOLO prediction layer is constructed by fusing more shallow features with deeper features. Moreover, the convolutional attention mechanism module CBAM is used in the neck network to improve the anti-background interference capability by strengthening the detection target features. In particular, a dataset of more than 17,000 sonar images containing three types of targets in a water tank experiment was built. Ultimately, compared with classic lightweight networks, the proposed network has the best performance in the dataset. Results illustrate the effectiveness of the proposed network, achieving a satisfactory balance between performance and consumption, which satisfies the requirements of high accuracy, small model size, and real-time.