Research on Feature Fusion Image Defogging Algorithms based on Transformers

Abstract

Under hazy conditions, airborne particulate matter and chemical substances reduce image contrast and blur details by absorbing and scattering light, severely compromising the accuracy of computer vision tasks such as autonomous driving and remote sensing monitoring. Addressing the limitations of traditional physical defogging models prone to estimation errors, the insufficient global feature modeling capabilities of mainstream convolutional neural networks (CNNs), and the high computational complexity and inadequate local detail capture of existing Transformer-based methods, this paper proposes a Transformer-based feature fusion image defogging algorithm. This algorithm employs a U-shaped encoder-decoder as its backbone network. It introduces a Hub-and-Spoke Multi-Head Attention (HSMHA) mechanism to replace traditional self-attention, significantly reducing computational overhead while preserving global context modeling capabilities. A Feature Refinement Block (FRB) is embedded within the feedforward neural network to enhance the recovery of image texture and detail information. A Multiscale Residual Enhancer (MRE) is constructed to effectively eliminate redundant high-frequency features and deepen learning of subtle feature variations. A contrastive regularization (CR) learning strategy is introduced, using blurred images as negative samples and clear images as positive samples to guide the model toward learning more discriminative feature representations, thereby enhancing the consistency between defogged images and their original clear counterparts. Experimental results on the SOTS-indoor and SOTS-outdoor synthetic datasets demonstrate that the proposed algorithm achieves optimal or suboptimal levels in both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Specifically, on the SOTS-indoor dataset, PSNR reaches 35.79dB and SSIM attains 0.984, and on the SOTS-outdoor dataset, PSNR was 34.31 dB and SSIM is 0.987. Qualitative results demonstrate the algorithm's superior performance in restoring image color fidelity, contrast, and detail integrity, effectively addressing issues such as incomplete defogging, color bias, or edge blurring present in existing methods.