SSS Underwater Target Image Samples Augmentation Based on the Cross-Domain Mapping Relationship of Images of the Same Physical Object

Side-scan sonar (SSS) image sample augmentation plays an important role in improving the effect of deep-learning-based underwater target detection. However, the existing sample augmentation methods for cross-domain conversion always result in weak representativeness of the augmented samples since the targets in the nondomain images are similar but not exactly the same as the actual underwater target to be detected. In this article, an augmentation method for SSS image samples of underwater targets based on the cross-domain mapping relationship of images of the same object is proposed. A physical model of the actual underwater target was first constructed using three-dimensional printing. A series of optical images and SSS images of underwater targets can be obtained by using the actual measurement of underwater targets under different conditions. To achieve the augmentation of SSS target samples, a single-cycle-consistency network structure with a channel and spatial attention and generative adversarial networks with least squares loss was designed for efficient and robust conversion of information between optical and SSS acoustic samples. To verify the effectiveness of the proposed method in generating high-quality samples, underwater targets were detected using the detection model trained by the generated samples. The experimental results revealed that the proposed method achieved impressive performance with a more than 5.8% improvement in average precision value for zero-sample underwater mine target detection and 4.3% for few-sample shipwreck target detection, compared with using only real SSS data.

Automated summary

This article proposes an augmentation method for SSS image samples of underwater targets based on the cross-domain mapping relationship of images of the same object. By constructing a physical model and obtaining a series of optical images and SSS images of underwater targets under different conditions, high-quality samples are generated and used to train a detection model. Experimental results show that this method achieves impressive performance with significant improvements in underwater mine target detection and shipwreck target detection compared to using only real SSS data.