Underwater Image Restoration for Seafloor Targets with Hybrid Attention Mechanisms and Conditional Generative Adversarial Network

Underwater image restoration is of great significance in underwater vision research. Affected by the underwater scene, the original underwater image usually has problems such as the color deviation, underexposure, noise, blur and less effective information. To address the above problems, we present an underwater image restoration method for seafloor targets with a hybrid attention mechanism-based conditional generative adversarial network. Firstly, we design a U-net generator network for encoding and decoding the input image. In the skip connections of U-net, we develop a hybrid attention mechanism module that includes spatial and channel attentions to enhance the depth of the network. Then, a multi-modal loss function is designed in the generator network, which takes into account the global content, color, local texture, image gradient, and style information. Qualitative and quantitative evaluations on the reconstructed images are conducted by comparing the proposed method with many previously published works. Experimental results demonstrate that the proposed method generates more visually appealing images and provides higher objective evaluation index scores. Furthermore, ablation experiments are implemented to verify the effectiveness of the proposed hybrid attention module. Noise robustness experiments show that the proposed method also has good denoising ability. Application testing experiments prove that the proposed method has a good application prospect in the seafloor target detection area.