基于特征解耦的无监督水下图像增强

Unsupervised Underwater Image Enhancement Based on Feature Disentanglement

水介质的吸收和散射特性致使水下图像存在不同类型的失真,严重影响后续处理的准确性和有效性。目前有监督学习的水下图像增强方法依靠合成的水下配对图像集进行训练,然而由于合成的数据可能无法准确地模拟水下成像的基本物理机制,所以监督学习的方法很难应用于实际的应用场景。该文提出一种基于特征解耦的无监督水下图像增强方法,一方面,考虑获取同一场景下的清晰-非清晰配对数据集难度大且成本高,提出采用循环生成对抗网络将水下图像增强问题转换成风格迁移问题,实现无监督学习;另一方面,结合特征解耦方法分别提取图像的风格特征和结构特征,保证增强前后图像的结构一致性。实验结果表明,该方法可以在非配对数据训练的情况下,能够有效恢复水下图像的颜色和纹理细节。

The absorption and scattering properties of the water medium cause different types of distortion in underwater images,which affects seriously the accuracy and effectiveness of subsequent processing.At present,underwater image enhancement methods with supervised learning rely on synthetic underwater paired image sets for training.However,the supervised learning methods are challenging to apply to practical application scenarios because the synthetic data may not accurately model the underlying physical mechanisms of underwater imaging.An unsupervised underwater image enhancement based on feature disentanglement is proposed.On the one hand,considering the difficulty and high cost of acquiring clear-unclear paired datasets in the same scene,a cycle generative adversarial network is employed to convert the underwater image enhancement problem into a style transfer problem to achieve unsupervised learning.On the other hand,the feature disentanglement method is combined to extract the style features and structure features separately to ensure the structural consistency of the images before and after enhancement.The experimental results show that the method can effectively recover the color and texture details of underwater images in the case of unpaired data training.