Weather Recognition based on Attention Image Search Method

Abstract

Weather monitoring plays a vital role in intelligent traffic transportation, and the improvement of weather recognition accuracy can effectively improve driving safety. At present, classification-based and segmentation-based algorithms for weather recognition have achieved good performance, but it is still full of challenges in real applications. On the one hand, the number of classes in public data sets is insufficient, which cannot identify the conditions such as stagnant water and debris flow. On the other hand, the current weather recognition methods have poor generalization ability, the model needs to be retrained when classes are changed. In this paper, we first propose a new multi-traffic weather (MTW) data set for weather recognition, it contains much richer classes. Then, a new weather recognition method based on attention image retrieval (AIR) is proposed to improve the performance of recognition. Compared with the previous methods, our method can obtain better generalization performance.