Channel and Spatial Transformer for Underwater Image Enhancement

Underwater images always suffered from significant degradation, such as color deviation, low contrast and blurred details. Underwater image enhancement is an essential step to improve the image quality. However, the existing underwater image enhancement results always have the problem of colour bias, and do not consider the non-uniform attenuation in different color channels. In this paper, a channel and spatial (CS) transformer module is proposed for the underwater image enhancement, where multi-dimensional attention is extracted from the deep features. The channel-level multiscale feature fusion (CMFT) module focuses on the channel with severe color attenuation. The spatial-level attention module(SAT) focuses on the degraded regions in spatial domain. Furthermore, a new loss function that combines perceptual losses in RGB, LAB and LCH colour spaces is proposed to correct color distoration and improve image details. A large number of experiments on existing datasets have verified the excellent performance of the proposed networks.