Parallel feature fusion framework for image dehazing driven by haze-detail collaboration

Image dehazing remains a highly ill-posed problem in low-level computer vision, primarily due to the complex coupling of spatially variant haze and high-frequency background details. Existing deep learning-based methods often struggle to balance the trade-off between aggressive haze removal and the preservation of fine textures, frequently resulting in color distortion, halos, or detail loss. To address these limitations and achieve high-fidelity restoration, this paper proposes a parallel feature fusion framework driven by a haze-detail collaboration mechanism. Specifically, the proposed framework adopts a dual-branch parallel architecture to disentangle the dehazing process. The upper branch functions as a haze layer extraction network. It employs a Res2Net-based encoder to capture multi-scale semantic features and integrates a novel deformable convolution-residual hybrid attention module. By dynamically adjusting the receptive fields, this module precisely characterizes non-uniform haze distributions and models long-range dependencies. Simultaneously, the lower branch serves as a detail compensation network, leveraging context detail information blocks with multi-scale dilated convolutions to aggregate contextual cues and reinforce the representation of high-frequency textural details. Subsequently, a fusion network performs adaptive feature integration, effectively merging the extracted haze features with the enhanced detail information to reconstruct the haze-free image. To ensure robust training, a dual-supervision mechanism is introduced, combining a feature regularization loss to align feature distributions in the latent space and a reconstruction loss to constrain pixel-level content fidelity. Extensive quantitative and qualitative experiments are conducted on both synthetic benchmarks and real-world datasets. The results demonstrated that the proposed algorithm delivered superior performance, achieving higher peak signal-to-noise ratio and structural similarity scores compared to state-of-the-art methods. Visual comparisons further confirmed that our method effectively removed dense haze while recovering vivid colors and sharp structural details without introducing artifacts.