摘要
Abstract
This research aims to propose an underwater image enhancement algorithm based on a self-attention model to address the issues of information loss,insufficient generalization ability,and cumbersome parameter adjustment in traditional underwater image enhancement algorithms.The aim is to improve the quality of underwater images in terms of brightness,color correction,and contrast,thereby better serving the fields of marine resource exploration and underwater exploration.Firstly,to solve the problems of poor feature extraction and detail information loss in the U-Net network,ST and MFFM modules are introduced.The ST module employs window-based multi-head self-attention(W-MSA)and shifted window-based multi-head self-attention(SW-MSA)mechanisms to effectively capture global image information and enhance local detail features.The MFFM module,through a combination of 1×1 and 3×3 convolutions and skip connections,improves the model's receptive field and inter-channel feature fusion capabilities.Secondly,feature map outputs are introduced into the discriminator to better process local detail information of the model.Finally,a multi-loss function fusion strategy,including content-based perceptual loss,adversarial loss,and L2 loss,is adopted to comprehensively evaluate and optimize the performance of the enhancement model.Experimental results show that the SM-GAN algorithm has achieved remarkable results in processing seabed images.Compared with the existing algorithms,the PSNR indicator is improved by 2.35%,UICQE by 4.95%and IE by 1.95%respectively.The brightness,color correction and contrast of the processed deep-sea images have been significantly improved,the image details are clearer,and the color shift problem has been effectively solved.关键词
图像增强/生成对抗网络/自注意力模型/深度学习Key words
image enhancement/generative adversarial network/gelf-attention model/geep learning分类
信息技术与安全科学
