Side-scan sonar underwater target segmentation using the BHP-UNet

Abstract Although target detection algorithms based on deep learning have achieved good results in the of side-scan sonar underwater targets, their false and missed rates are high for multiple densely arranged overlapping targets. To address this problem, a segmentation model blended hybrid dilated convolution pyramid split attention UNet (BHP-UNet) algorithm is proposed paper. First, module adopted to improve ability learn semantics shallow features while improving receptive field. Second, introduced establish long-term dependency between global local information processing multi-scale spatial features. Three sets experimental show that BHP-UNet paper has better performance than conventional fully convolutional network, UNet, DeepLabv3+ models, it able segment dense targets certain extent. The will significance as guide practical applications.