The Method of UAV Image Object Detection under Foggy Weather by Style Transfer

Abstract

Under the bad weather conditions such as foggy or rainy weather, the details of UAV images are seriously lost. Consequently, directly detecting without dehazing may lead to object missing and wrong inspection. Although the method of dehazing firstly and then detecting can significantly improve the detection performance of UAV images under misty weather conditions, it can not solve the problem that object detection model relies too much on the texture and surface information of the targets in the image which are used to classify, and thus leads to a poor detection performance on heavy fog images. In response to this problem, a method to detect after style transfer instead of haze removal is proposed to accomplish object detection of the UAV images under dense fog weather conditions. The method can change the style of the image while keep the content in image unchanged, making the object detection model transform from learning the object texture and surface information to learning the object contour, which is a more difficult task. The experiment is carried out on the UAV dataset Visdrone2019. Firstly, the UAV images are divided into fog-free, mist and dense fog images by using the differences in gradient features, dark channel features and wavelet features of fog images. Secondly, the processing methods of dehazing and style transfer are adopted respectively according to the concentration of fog. The experimental results show that the proposed method can further reduce the impact of light and noise, and can significantly improve the object detection performance of UAV dense fog images. Moreover, combining with the method of dehazing firstly and then detecting, the proposed method can adaptively complete the object detection task of UAV fog images.

Keywords

• |uav image|fog image|style transfer|target detection|image classification