An effective automatic object detection algorithm for continuous sonar image sequences

Object detection of continuous sonar image sequences has become an efficient way for underwater environment exploration. However, the task always suffers from the influence of th e complex underwater environment. In particular, the existing algorithms mainly focus on image-based detection and can not balance the detection speed and accuracy in continuous sonar image sequences. To solve the problem, this paper proposes a novel automatic detection algorithm based on deep learning for continuous sonar image sequences. Firstly, the convLSTM (convolution Long Short-Term Memory) is improved to fuse sonar features obtained from the cross-detection model, which are from three aspects: 1) The original convolution is replaced by depthwise separable convolution; 2) The original network input is divided into G groups and processed by group convolution; 3) A connection layer between the Bottleneck convolution layer and output is added to further capture feature information between frames. Then, to fully extract sonar features, a cross-detection network is established by fusing two different feature extraction networks MobileNetV3-Large and MobileNetV3-Small. Finally, we combine thecross-detection network with the improved convLSTM to establish the whole model, which can fully extract and utilize temporal information in continuous sonar image sequences. The experimental results show that the proposed model has effectively improved the detection speed in sonar image sequences at 150 FPS, simultaneously keeping an 85.8% mAP.