Dynamic Cooperative Attention Based De-Fogging Networks

To address the limitations of convolutional neural networks in deep feature extraction, this study proposes a dynamic deformable convolution with a four-way parallel structure in the encoder. The method adaptively adjusts the convolutional kernel sampling positions through a learnable offset prediction network, thereby enhancing the model's ability to perceive non-uniform fog distribution. Furthermore, to overcome the constraints of existing attention mechanisms in cross-scale feature modeling, a dual-path collaborative attention feature transformation layer is introduced. This layer improves long-range dependency modeling through interactive information transfer between structural perception and semantic enhancement pathways. Building upon these innovations, this study develops the Dynamic Collaborative Attention Enhancement Network (DCAE-Net). The network achieves precise modeling of haze degradation by leveraging an encoderdecoder architecture and a synergistic design of multi-level feature interaction and dynamic perception units. The key contribution lies in integrating dynamic deformable convolution, cross-modal attention mechanisms, and adaptive fusion strategies to establish a hierarchical processing system with spatial perception capabilities. Experimental results demonstrate that the proposed network outperforms existing methods in preserving image details and color fidelity.