Single Image Dehazing via Multi-Scale Large Kernel Convolutional Neural Networks

Large-kernel convolutional neural networks (CNNs) have recently achieved remarkable performance comparable to Visual Transformers(ViTs) in high-level vision tasks. However, there are two critical drawbacks hindering its widespread applications in image dehazing. 1) Most large-kernel designs focus expanding the kernel size even further to model stronger long-range dependencies, but this approach brings a substantial amount of computational overhead. 2) As the kernel size increases, the network tends to focus more on the shape of the object over its texture, potentially affecting the details of the recovered image. To overcome these issues, we propose an effective multi-scale large separable kernel attention module (MLSKA) that can simultaneously build long-range and local dependencies in a cost-effective manner to facilitate high-quality image reconstruction. Specifically, MLSKA combines an efficient convolutional decomposition design with multi-scale learning, realizing multi-scale receptive fields while significantly reducing the computational cost and parameters of large-kernel convolution. In addition, we introduce a deformable attention feed-forward network (DAFN) to aggregate contextual information. In DAFN, a novel deformable attention gate is designed to provide holistic attention to the feed-forward network (FFN), thereby improving its utilization of critical features. Integrating these two designs into a U-shaped backbone, the proposed multi-scale large-kernel network (MLANet) outperforms state-of-the-art methods on several dehazing benchmarks, achieving the best parameter-performance trade-off.