DCPFNet: dark channel prior guided frequency network for image dehazing

Image dehazing is a vital task in computer vision, enhancing the robustness and perceptual capability of outdoor vision systems. Recent studies increasingly employ frequency-domain analysis for its advantages in global contrast restoration and high-frequency detail recovery. However, existing frequency-based methods often treat spectral enhancement and physical degradation as independent processes, lacking effective physical constraints during frequency modulation. Given the spatially non-uniform haze distribution, frequency enhancement should be spatially adaptive: dense haze regions need stronger frequency responses for structural recovery, whereas clearer areas should preserve more spatial-domain details. To address this, we propose a dark channel prior guided frequency network (DCPFNet) for image dehazing. An enhanced dark channel prior module first extracts spatially varying haze concentration distributions, refined by a learnable network for improved accuracy and robustness. A spectral gated network then performs amplitude–phase decoupling and gated enhancement in the frequency domain, mitigating global contrast degradation and high-frequency loss. The key innovation is the physics-guided adaptive spectral processing mechanism, which embeds dark channel priors into the frequency pipeline, enabling adaptive spectral modulation according to local haze density for region-specific degradation compensation. The overall lightweight dual-stage cascaded architecture employs cross-stage feature fusion for progressively refined reconstruction. Extensive experiments on synthetic and real-world datasets demonstrate that DCPFNet surpasses state-of-the-art methods, especially under heavy haze and complex illumination. Ablation studies further confirm the synergy between physical-prior guidance and frequency-domain learning, emphasizing that integrating physical knowledge into frequency-domain processing is crucial for high-quality image dehazing.