Multi-class weather classification based on multi-feature weighted fusion method

Abstract

Weather classification, which aims to prevent extreme weather from affecting people's life and property, has recently become a significant interest in image classification. The traditional weather recognition methods always adopt expensive sensors and human vision as auxiliary. However, they perform not very well. Although the Convolutional Neural Networks (CNNs) have shown the state of the art performance on image classification recently, for weather classification, it is a challenging task because it relies on weather-sensitive cues such as the various illumination, contrast and the clutter background of the weather images. To address this issue, a method named multi-feature weighted fusion is presented in this paper. Firstly, we extract the well-chosen weather-specific features such as haze, contrast, brightness, etc., and then fuse these features with data-driven CNNs feature to form high dimensional vectors. Moreover, to increase the recognition performance, various weights are adopted for weather-specific features and CNNs feature to learn an adaptive five-class weather conditions classifier. The extensive experimental results demonstrate that the proposed method obtains better performance than the method using CNNs feature alone.