HU-Net: A Hybrid U-Net for Single Image Dehazing Using Composite Attention Module

Image dehazing plays a crucial role in computer vision applications encompassing categorization, object identification, and picture restoration in adverse weather circumstances. Dehazing is a problem that lacks a clear solution, and to address it, many methods such as transformers and convolutional neural networks (CNNs) have been suggested. The transformer-based and CNN-based methods employ distinct approaches to extract features and yield disparate representations of features for a particular image. The performance of transformer and CNN-based approaches vary due to differences in their properties. This study introduces a composite attention block for Single-image dehazing, which aims to integrate the performance of both approaches. The attention block is a deep learning method that selectively focuses on specific areas of the input to enhance accuracy and computational efficiency. In this case, we employ a composite attention block that combines CNN-based and vision-based attention models to extract features. The entirety of our network is referred to as Hybrid U-Net (HU-Net), and it is constructed using the U-Net design. The HU-Net consists of an encoder and decoder, both equipped with skip connections. Additionally, it incorporates a fusion model for feature fusion. The HU-Net consists of blocks that utilize a composite attention module instead of the Multi-Head Attention Network to extract features. The HU-Net is trained using a hybrid loss function (HLF) on three different picture datasets: O-Hazy, I-Hazy, and NH-Hazy. The suggested network demonstrates superior performance concerning PSNR and SSIM when compared to other established techniques.