Domain Adaptive Multi-task Learning for Complex Weather Images
Han Hui Li · 2024
Abstract: Deep convolutional neural network-based models are introduced to improve the performance of computer vision tasks. Although these models achieve good accuracy under good weather conditions, images acquired under harsh weather conditions, such as rain, snow, fog, etc., can be degraded and blurred, making it difficult to extract effective features. Therefore, the accuracy of these models can be greatly reduced when working under these conditions. To address the above issues, we design a multi-task learning algorithm based on domain adaptation method, using a multi-phase teacher-student model to achieve transfer learning from good weather to complex weather. The multi-task learning part simultaneously implements object detection and semantic segmentation tasks, improving model training and data usage efficiency. In addition, a complex weather simulation algorithm is introduced to simulate complex weather, making the model capable of learning multiple weather conditions simultaneously and more robust. Experiments show that this method performs better than existing methods.