An Underwater Sonar Small Sample Image Transformation for Feature Domain Optimization

Underwater Image transformation uses deep learning to generate images with the target domain and the source domain. It can quickly and accurately solve transformation problems such as image blurring for underwater images. To solve the problem of the poor transformation effect caused by the limited number of sonar images, we propose an underwater small sample image transformation (SSIT) for the feature domain with representation enhancement. We use Canny and histogram equalization to extract the details of images. Then, a channel attention mechanism with sub-pixel convolution is applied to improve the feature extraction ability and enhance image representation. We adopt a noise addition strategy and an image reconstruction strategy to scale the loss of cycle consistency. The weakened cyclic semantic consistency is applied to improve the quality of transformed images. Due to the different image datasets having different data characteristics, the underwater sonar small sample dataset is selected for image transformation with Generative Adversarial Networks (GAN). We choose FUNIT as the baseline, and do the ablation experiment to demonstrate the validity of the model. Comparing with different situations, the experimental results show that the image representation enhancement constructed in this paper can effectively enhance the details of images.