Abstract: Aiming at the problems of detail blur, low contrast and blue-green skew in underwater degraded images, an underwater image enhancement algorithm is proposed based on multi-scale feature fusion to generate antagonistic network. Based on the generated adversarial network, this algorithm combines the traditional white balance algorithm and multi-scale enhancement network to enhance underwater degraded images. Firstly, an white balance algorithm with improved channel compensation is used to correct the color bias of underwater degraded images, and the features of the corrected images are extracted by convolutional neural network. Then, the multi-scale features of the image are extracted, and the local features of each layer are enhanced to capture the global features of semantic information by combining with the proposed residual dense blocks, which are fused with the features of the color correction images. Finally, the fusion features are reconstructed into clear images by the reconstruction module, and the details of the images are recovered. The experimental results show that the enhanced underwater image with this algorithm has better defogging effect and more real color, and effectively improves the problem of color bias and blur of underwater image. The experimental results are superior to the comparison algorithm in both subjective and objective indicators.

Key words: underwater image enhancement, generative adversarial network, multiscale, feature fusion

摘要： 针对水下退化图像细节模糊、对比度低和蓝绿色偏问题，提出了一种基于多尺度特征融合生成对抗网络的水下图像增强算法。算法以生成对抗网络为基本框架，结合传统白平衡算法和多尺度增强网络实现对水下退化图像的增强。通过改进的通道补偿白平衡算法矫正蓝绿色偏，并以卷积神经网络提取偏色校正后图像的特征；提取图像多尺度特征，结合提出的残差密集块将每一层的局部特征增强为捕获语义信息的全局特征，并与偏色校正图像的特征相融合；通过重建模块将融合特征重建为清晰图像，恢复图像的细节信息。实验结果表明，该算法增强的水下图像去雾效果较好且颜色更真实，有效改善了水下图像色偏和模糊的问题，在主观指标和客观指标上的实验结果均优于对比算法。

关键词: 水下图像增强, 生成对抗网络, 多尺度, 特征融合