Cascaded Multi-level Features Fusion Network using Collaborative Training for Snow Removal

Visible atmospheric particles (such as raindrops, haze particles, and snow particles) could mask the image pixels in the form of some noise, which makes many computer vision tasks difficult to implement. At present, most existing neural network models for snow removal exploit the principle of snow generations, and they remove the snow particles through extracting the snow masks and chromatic aberration maps from the snowy images. Typically, the original image features and the blurred image features are not jointly considered. In this paper, according to the transparency and size of snow particles, also the residual masks between snowy images and snow-free images, we propose a Cascaded Snow Removal Network (CSR-Net). Moreover, the proposed model adopts the collaborative training strategy to accelerate the training process, effectively harnessing the computational resources across multiple devices. Besides, the model parallelism ensures the collaborative training among devices, thereby improving the overall model performance. Specifically, our proposed method can effectively enhance the Peak Signal to Noise Ratio (PSNR) and Structural SIMilarity (SSIM).