Learning enriched channel interactions for image dehazing and beyond

Abstract

Atmospheric haze degrades image clarity and impairs the performance of downstream computer visions tasks. Convolutional neural networks have demonstrated strong dehazing capabilities by exploiting neighborhood spatial patterns, while Vision Transformers excel at modeling long-range dependencies. However, existing methods suffers two challenges. First, inadequate modeling of inter-channel correlations leads to wavelength-dependent color distortions. Second, insufficient preservation of frequency-specific components results in blurred textures under non-uniform haze distributions. To tackle these limitations, we present the Dual-Domain Channel Attention Network (DDCA-Net), which integrates Spatial Channel Attention (SCA) and Frequency Channel Attention (FCA). The SCA module explicitly models spatial inter-channel dependencies to correct color imbalances, and the FCA module employs a multi-branch frequency decomposition mechanism to selectively restore high-frequency details attenuated by haze. This dual domain approach enables the precise reconstruction of fine-grained structures while enhancing overall image clarity. Extensive evaluations of nine benchmark datasets demonstrate consistent improvements over state-of-the-art methods. In particular, DDCA-Net achieves PSNR gains of 0.32 dB on RESIDE Indoor, 0.88 dB on SateHaze1K, and 1.79 dB on LOL-v2. Furthermore, our model yields significant boosts in downstream object detection and segmentation, confirming its practical utility. The code is available at https://github.com/hafeezbabar/DDCA-Net.