<p>Image dehazing, a fundamental challenge in low-level vision, plays a critical role in restoring image clarity compromised by atmospheric phenomena. In recent years, Transformers have advanced various dehazing tasks via robust long-range dependency modeling, but self-attention’s quadratic time complexity hinders their practical deployment. We propose Bandwise Rectangle Attention Fusion Network (BRFNet), a novel CNN architecture integrating three key innovations: the Rectangle Attention Module (RAM) with linear complexity, which outperforms prior strip-based attention in efficiency by capturing row/column-wise pixel context via convolution-learned weights and enhances multi-scale learning through varied strip dimensions; the Bandwise Attention Module (BAM), a transformation-free component for lightweight spectral refinement that segregates and adjusts frequency components to boost feature capture; the Fusion Attention Block (FAB), a parallel structure tailored to haze physics that extracts location-specific local and shared global information simultaneously, effectively handling non-uniform haze distributions (unlike single transformer feed-forward modules). Unlike prior works focusing on isolated aspects, BRFNet synergizes these elements into a unified, efficient, and highly effective framework. Experiments on three benchmark datasets confirm BRFNet’s leading performance.</p>