Unsupervised Domain Adaptive Learning for Image Desnowing with Real-World Data

Snow images usually contain snow grains, snow streaks, and mist, which greatly affect the visibility of images. Currently, supervised learning with synthetic data often faces limitations when it comes to handling real-world snow images. To address this crucial issue, this work proposes an unsupervised domain adaptation image snow removal framework. The framework improves the performance on real-world images by learning a domain classifier in adversarial training manner. Additionally, considering the diversity of snowflake shapes and sizes in real-world snow images, we design a multiple-kernel dilated convolution module. Extensive experiments on three representative datasets have validated that our model can achieve better results than existing desnowing methods. More importantly, experiments on real datasets show that the proposed method obtains state-of-the-art performance in real-world desnowing.

1. INTRODUCTION

Snow grains often occlude or blur the important information of the images captured outdoors. Thus the snow removal task for an image or a video is useful and necessary, which can be served as an important pre-processing step for many high-level vision tasks, such as object detection [1], pedestrian detection [2], and video surveillance [3]. Snow removal is an important task that has received increasing attention in the field of computer vision.