Research on Underwater Image Restoration Algorithm Based on Generative Adversarial Network

Underwater optical imaging is prone to image distortion, blurring and other problems arising from thermal disturbance. To tackle down the above-mentioned problems, underwater images affected by thermal disturbance are analyzed in this study from two perspectives (i.e., image gray histogram distribution and similarity assessment). Subsequently, a deep learning-based underwater thermal disturbance image restoration algorithm is developed based on generative adversarial networks. The U-Net structure is employed in the generator network, and MultiRes blocks are adopted in the contracting and expanding paths for extracting image feature information on different scales maximally. Moreover, a Bottleneck Attention Module (BAM) is introduced in the contracting path to place a full focus on the distorted image information. High-precision image restoration is achieved using the Wasserstein adversarial loss function in the adversarial process between the generator and discriminator networks. To validate the algorithm, an underwater thermal imaging experimental platform is independently built, and a dataset is established. As indicated by the experimental results, the proposed algorithm is capable of effectively restoring images while preserving key edge features, such that a Structural Similarity (SSIM) index of 0.7749 and a Peak Signal to Noise Ratio (PSNR) index of 21.28 are generated. This study can lay a solid foundation for further processing in underwater image restoration.