基于多尺度残差注意力网络的水下图像增强

Underwater image enhancement based on multiscale residual attention networks

针对水下图像由水的散射、吸收引起的色偏、色弱、信息丢失问题,提出了一种基于多尺度残差注意力网络的水下图像增强算法。该网络引入了改进的UNet3+-Avg结构与注意力机制,设计出多尺度密集特征提取模块与残差注意力恢复模块,以及由Charbonnier损失和边缘损失相结合的联合损失函数,使该网络得以学习到多个尺度的丰富特征,在改善图像色彩的同时又可保留大量的物体边缘信息。增强后图像的平均峰值信噪比(PSNR)达到23.63 dB、结构相似度(SSIM)达到0.93。与其他水下图像增强网络的对比实验结果表明,由该网络所增强的图像在主观感受与客观评价上都取得了显著的效果。

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.