Knowledge Transfer via Leveraging Teacher-Student Network with Visual Attention to Enhance Atmospheric Sand Image Restoration

Deep learning-based methods have recently shown superiority in sand dust image restoration tasks. However, most existing learning-based methods do not pay much attention to the gap between the synthetic image and the real image, making the obtained recovery effect difficult to generalize on the real image. To alleviate this problem, we propose a novel Teacher-Student network (TS-Net) for real-world sand dust image restoration that maximizes the transfer of knowledge learned in the Teacher Network. Thereby improving the recovery ability of the Student Network on real sand dust images. Firstly, the Teacher Network extracts the latent feature information of dust images. Furthermore, use transfer learning to transfer the knowledge learned by the Teacher Network to the faster and more efficient Student Network. We present a Visual Attention (VA) mechanism to couple multi-scale features with contextual features with large kernel convolutions to fully utilize these complementary features to correct the color of sand dust images. Due to the large kernel convolution and feature fusion strategy, the fused features can correct the color cast the problem of dust color, which is very important for high-level tasks. Extensive experiments on the synthetic and real datasets demonstrate that the proposed method significantly improves over the state-of-the-art methods.