TransID: Image Dehazing Based on TransUNet

Because UNet works well in learning damaged picture restoration from large-scale data, these models are commonly utilized in image dehazing applications. Transformer, another type of network design, has recently proven to be effective in natural language and complex visual tasks. Transformer models improve convolutional neural network limitations, but they fall short of UNet in terms of texture and detail repair. In this paper, we present an effective and efficient hierarchical encoder-decoder network architecture based on TransUNet. Our network has two core designs to make it suitable for this task. The first essential component is a TransUNet-based multi-scale dense feature fusion network, where we merge global self-attention from Transformer and multi-scale context data from UNet. The second essential component is our investigation of the Context Information Compensation Module to add-on data from the encoder to the decoder. With the use of these two designs, our strategy keeps the benefits of the conventional UNet network while enhancing the network with the new transformer's superior performance in terms of global self-attention. The results of our experiments show that both the quantitative and qualitative evaluations are better than the current methods.