Generating Underwater Images by GANs and Similarity Measurement

Abstract

The complexity of the underwater environment makes it difficult to capture underwater images, therefore can not provide a large amount of training data required for image processing with deep learning. Traditional color transfer methods can only map the color distribution of source image to that of target image one by one, without considering the complexity of the underwater imaging environment, which results in the disadvantages of effectiveness and performance. In this paper, our proposed method, based on GANs(Generative Adversarial Networks), is superior in its consideration of taking underwater imaging model to train the generative network, the air images are converted into underwater images in batches, so as to achieve the purpose of data augment. Additionally, we describe several implementation details to improve the performance of GANs. Finally, we first utilize SIMILATION, SSIM(Structural Similarity Index) and MS-SSIM(Multi-Scale-Structural Similarity Index) to compute the color and structure similarity level for the purpose of evaluating the quality of generated samples. The results indicate that the proposed method performs well in data augment.