Underwater Image Enhancement Network Based on Swin-Convs Transformer with Guidance Embedding

Abstract

Underwater image enhancement (UIE) approaches struggle with the problem of dataset collection and attempt to restore underwater photos that have been damaged by light absorption and dispersion. In the paper, we propose to combine depth and more a priori informations to generate different kinds of high quality datasets, and then propose an underwater enhancement algorithm that improves the Transformer with a priori information. Firstly, a novel medium transmission guidance position embedding model is well-designed to provide the Transformer with relative position information, color bias information and global features of the feature map. Second, the restored convolution can grab more local attention given Swin Transformer's shortcomings in this area. Therefore, we propose an ingenious way to fuse convolution and core attention mechanisms to capture more enough local attention and adopt a new converter encoder that uses an on-chip converter block to enhance on-chip attention both in channel-wise and spatial-wise, thus effectively eliminating weather degradation at details, which is enhanced in terms of channel and spatial attention of Swin Transformer. The experimental results show that the enhanced underwater image PSNR index and SSIM index of our proposal are higher than those of other competitors. The subjective quality is also significantly improved, and the enhanced images with rich color and high definition are produced.