ABSTRACT Single‐image dehazing remains a challenging problem due to non‐uniform haze distribution, which often leads to degraded visual quality and incomplete reconstruction in existing deep learning approaches. We propose RSDhazer, a residual shallow‐deep dehazer network designed to address these challenges through complementary feature extraction pathways. The architecture integrates three key modules: multi‐activated feature extraction for enhanced feature diversity through parallel activation functions, cascaded residual block for expanded receptive fields through hierarchical feature propagation and residual attention block for spatially adaptive feature refinement. By combining shallow and deep representations, RSDhazer effectively models complex haze degradations while preserving critical image attributes, including colour, texture details and edge sharpness. Extensive experiments on both synthetic benchmarks (SOTS, FRIDA) and real‐world datasets (O‐Haze, I‐Haze, N‐Haze) demonstrate that RSDhazer achieves state‐of‐the‐art performance with high PSNR and SSIM scores, validating its robustness and generalisation capability across diverse haze conditions.
Zahra et al. (Thu,) studied this question.