DMNet: Image dehazing via Dual-Domain Modulation

Abstract

Feature representation of hazy images is directly related to the performance of dehazing models. However, existing approaches often struggle to jointly model spatial-frequency domain characteristics for characterization of heterogeneous haze distribution. Therefore, to address these challenges, this work presents an efficient Dual-Domain Modulation Network (DMNet) for image dehazing, which enhances the representation of uneven haze features and the global feature perception by utilizing the deformable convolution and the amplitude-phase guidance strategy. For one thing, since fixed-size convolutions are inadequate for multi-scale feature extraction and inter-channel interactions, we propose the Deformable Convolutional Operator (DCM) based on the spatial non-uniform strategy of channel interactions. Through orthogonal spatial feature aggregation mechanism, the DCM effectively aggregates spatial context information to handle non-uniform haze distribution and reconstruct fine texture details in heavily hazed regions. For another, the amplitude-centric reconstruction paradigm fails to accurately represent the nonlinear mapping between hazy and clear images in the frequency domain and neglects the importance of phase structural feature in image reconstruction. Therefore, we propose the Amplitude-Phase Guidance Module (APGM) to effectively extract global features through implementing low-pass filtering on the amplitude component and high-pass filtering on the phase component. Ultimately, by combining DCM and APGM, we propose the Dual-Domain Modulation Module (DM), which serves as the core component of DMNet to overcome the hurdles faced in achieving fusion between spatial and frequency domains. Extensive experiments demonstrate that DMNet performs favorably against the state-of-the-art (SOTA) approaches, achieving the PSNR over 41.77 dB with only 3.94M parameters.