Weather-degraded image semantic segmentation with multi-task knowledge distillation

The semantic segmentation of degraded image in adverse weather is of great importance for the navigation system of autonomous driving. However, weather-degraded images increase the difficulty of semantic segmentation as well as decrease the accuracy. It is natural to integrate image enhancement into degraded image semantic segmentation to improve the accuracy, which is computation intensive and time consuming. To meet the challenge, we propose a fast degraded image semantic segmentation with Multi-Task Knowledge Distillation called MTKD. The proposed MTKD method encourages image enhancement and semantic segmentation networks to learn from each other to make full use of the correlation between two tasks. Additionally, we propose shift operator to realize a lightweight model design. Extensive experiments demonstrate that the proposed MTKD outperforms state-of-the-art methods not only with better semantic segmentation performance but also with higher speed in weather-degraded images, which achieves 0.038 s in semantic segmentation for a 2048 × 1024 image.