多输入融合对抗网络的水下图像增强

Multi-input fusion adversarial network for underwater image enhancement

针对水下图像出现对比度低、颜色偏差和细节模糊等问题,提出了多输入融合对抗网络进行水下图像增强。该方法主要特点是生成网络采用编码解码结构,通过卷积层滤除噪声,利用反卷积层恢复丢失的细节并逐像素进行细化图像。首先,对原始图像进行预处理,得到颜色校正和对比度增强两种类型图像。其次,利用生成网络学习两种增强图像与原始图像之间差异的置信度图。然后,为减少在生成网络学习过程中两种增强算法引入的伪影和细节模糊,添加了纹理提取单元对两种增强图像进行纹理特征提取,并将提取的纹理特征与对应的置信度图进行融合。最后,通过构建多个损失函数,反复训练对抗网络,得到增强的水下图像。实验结果表明,增强的水下图像色彩鲜明并且对比度提升,评价指标UCIQE均值为0.6399,NIQE均值为3.7273。相比于其他算法有显著优势,证明了该算法的良好效果。

For underwater image of low contrast,color deviation and blurred details and other issues,the multiinput fusion adversarial networks was proposed to enhance underwater images.The main feature of this method was that the generative network used encoding and decoding structure,filtering noise through convolution layer,recovering lost details through deconvolution layer and refining the image pixel by pixel.Firstly,the original image was preprocessed to obtain two types of images:color correction and contrast enhancement.Secondly,the confidence graph of the difference between the two enhanced images and the original image was learned by using the generated network.Then,in order to reduce artifacts and details blur introduced by the two enhancement algorithms in the process of generating network learning,the texture extraction unit was added to extract texture features from the two enhanced images,and the extracted texture features were fused with the corresponding confidence map.Finally,the enhanced underwater image was obtained by constructing multiple loss functions and training the adversarial network repeatedly.The experimental results show that the enhanced underwater image has bright color and improved contrast,the average value of UCIQE and NIQE is 0.6399 and 3.7273 respectively.Compared with other algorithms,the algorithm has significant advantages and proves its good effect.