Underwater image enhancement based on multiscale residual attention networks

An underwater image enhancement algorithm based on multi-scale residual attention network was proposed for the problems of color shift, color fading and information loss of underwater images caused by water scattering and absorption. An improved UNet3+-Avg structure and attention mechanism was introduced by the network, and the multi-scale dense feature extraction module as well as the residual attention recovery module were designed. In addition, a joint loss function combining Charbonnier loss and edge loss enabled the network to learn rich features at multiple scales, improving the image color while retaining a large amount of object edge information. The average peak signal-to-noise ratio (PSNR) of the enhanced images reaches 23.63 dB and the structural similarity ratio (SSIM) reaches 0.93. Experimental results with other underwater image enhancement networks show that the images enhanced by this network achieve significant results in both subjective perception and objective evaluation.