Real-time weather image classification with SVM

Abstract

Accurate and real-time weather image recognition is important for system stability, safety decisions, and resource scheduling. For example, in scenarios such as intelligent transportation systems, agricultural automation, and driverless vehicles, the system's accurate perception of current weather conditions directly affects its operating efficiency and safety. Traditional machine learning methods, such as SVM, often have low accuracy. In recent years, deep learning methods such as Convolutional Neural Networks (CNN) have shown high accuracy in image classification and weather recognition. But these methods usually need a lot of computing power, which limits their use in real-time systems. This leads us to explore more lightweight and computationally efficient traditional methods for weather image classification. To solve the problems, this thesis proposes an improved weather classification method based on SVM. This method is different from the traditional classification method based on a single feature, this thesis combines multiple image features to improve both accuracy and speed. We extract several features from raw weather images, including brightness, saturation, noise level, blur metric, edge strength, local binary pattern (LBP), and color histogram. These features form a feature vector used to classify images into four weather types: clear, low light, rainy, and haze. The results show that this multi-feature SVM method performs better than the traditional SVM on several public datasets. It is also much faster than CNN-based methods. This makes it more suitable for real-time applications.