Research on Adaptive Clarification Method for Complex Weather Images in Traffic Scenes
Xiaolin Shi, Huiqin Zhu, Xinqian Yang · 2023
Aiming at the problem of single application scenario of different weather image clarification algorithms in transportation scenarios, an adaptive clarification method for complex weather images is designed. Firstly, different types of complex weather images are collected, and a multi-category weather image dataset FRSO containing fog, rain, snow and other (sunny and cloudy) weather types is obtained by manual screening. Then, a Multi-scale Texture Attention (MTA) module is proposed based on the unique texture structure of the weather images and combined with Transfer Learning (TL), a lightweight weather recognition model WeatherNet-MTA-TL is constructed; Finally, the corresponding clarity algorithm is invoked for processing to realize the function of adaptive defogging, deraining and desnowing of complex weather images. The experimental results show that the weather recognition accuracy of WeatherNet-MTA-TL reaches 97.11%, which is higher than that of MobileNetV2-TL and EfficientNet-TL models by 1.44% and 0.67% respectively, and lower than that of ResNet101-TL model by 0.45%, and the number of its parameters is only 5.42M, which accounts for the percentage of 93.00%, 69.58% and 12.17% of MobileNetV2-T, EfficientNet-TL and ResNet101-TL models respectively, which has a smaller number of parameters while maintaining a higher recognition accuracy. The method proposed in this paper is able to clarify weather images in real time on the basis of automatic recognition of weather images with high accuracy, which meets the practical application requirements of vehicle detection, lane line detection, and other transportation scenarios.