Improved Swin Transform-based generative adversarial network underwater image enhancement model

Along with continuous deepening of ocean exploration, the underwater image enhancement and recovery technology is paid attention to. The invention provides a generative adversarial network underwater image enhancement model (SwinGAN) based on an improved Swin Transform, aiming at the problems of low contrast ratio, large noise, color deviation and the like of an underwater image. The method comprises the following steps of: performing bicubic interpolation preprocessing on an underwater image before inputting the underwater image into a generator, segmenting an input feature map into a plurality of non-overlapped local windows on a bottleneck layer, and enhancing local attention while enhancing global information capture and distance dependency relationship by using a double-channel window multi-head self-attention mechanism. And finally, recombining a plurality of windows in a decoder to form a feature map with an original size, and adopting an improved Markov discriminator as a discrimination network in the