Abstract:

In view of the scarcity, difficulty and high cost of side-scan sonar underwater target images, and the poor performance of deep-learning based target detection model, combined with the abundant target data set in optical domain, a sample augmentation method for underwater target side-scan sonar images based on channel and spatial attention(CSA) module and least squares generative adversarial networks(LSGAN) and cycle generative adversarial networks(CycleGAN) is propesed. Firstly, inspired by CycleGAN, a single cycle network structure based on cycle consistency is designed to ensure the training efficiency of the model. Then, the CSA module is integrated into the generator to reduce information dispersion while enhancing cross-latitude interaction. Finally, a loss function based on LSGAN is designed to improve the quality of the generated image while improving the training stability. Experiments are carried out on ship optical domain data set and side-scan sonar shipwreck data set. The results show that the proposed method achieves efficient and robust conversion of information between optical and side-scan sonar samples and augmentation of a large number of side-scan sonar target samples. At the same time, the underwater target detection is carried out based on the detection model generated after sample training in this paper. The results show that the average precision value of the model after training with sample augmentation data in this paper reachs 84.71% in detecting shipwreck targets with few samples, which proves that the method in this paper achieves high-quality amplification of highly representative underwater target samples with zero samples and small samples. It also provides a new way to construct high-performance underwater target detection model.

Key words: sample augmentation, side-scan sonar, cycle generative adversarial networks (CycleGAN), channel and spatial attention (CSA) module, least squares generative adversarial networks (LSGAN)