TEGAN: Transformer Embedded Generative Adversarial Network for Underwater Image Enhancement

Underwater robots are widely used in underwater missions. However, due to complex scenes, it is difficult to obtain high-quality underwater images, which usually suffer from severe distortions such as low visibility, blurred edges, and color cast. In this paper, a Transformer embedded generative adversarial network for underwater image enhancement is presented. We propose a window-based dual local enhancement block to compensate for the Transformer’s shortcomings in extracting local features and improving image clarity. Convolutional neural network is deployed in sequential and parallel modes for local enhancement. Second, for generator construction, a fusion scheme combining convolutional neural network and Transformer block in units is designed. We exploit a self-attention mechanism to extract long-distance dependencies and fully extract the original features at the initial stage to enhance the image details. Meanwhile, global information is captured through the bottleneck for color correction. Convolutional neural network, which is good at extracting local features, is introduced in Encoder/Decoder units for multiscale feature extraction and reconstruction to effectively reduce edge blurring. Finally, a Transformer embedded generative adversarial network with a two-branch discriminator is established to generate more realistic colors while preserving the image content. Comparative experimental results show that our method can achieve superior results to the state-of-the-art approaches on both paired and unpaired datasets. Excellent learning and generalization ability make it outperform others in subjective perception and overall performance evaluated by image quality metrics. In addition, the enhancement results also show the significant improvement it brings in the downstream visual application tasks.