Vehicle logo recognition based on a weighted spatial pyramid framework

Yuanchang Ou, Huicheng Zheng, Shuyue Chen, Jiangtao Chen

Abstract: This paper proposes an effective vehicle logo recognition framework, which is robust when the logos are only roughly located but not well segmented. Regions of interest (ROI) are first detected by using an AdaBoost-based detector. The detector is tuned to have a low false negative rate so as to guarantee coverage of the vehicle logo as much as possible. A weighted spatial pyramid framework is introduced to extract feature vectors from these ROIs. In this framework, we consider the union of ROIs instead of processing the ROIs individually for robustness and efficiency. Dense SIFT descriptors are extracted from the ROIs for robust description of the image. The scale-invariant feature transform (SIFT) descriptors are weighted based on the corresponding ROIs, highlighting locations with high confidence. The spatial pyramid scheme is then implemented to exploit the spatial distribution of local features. Finally, we apply a linear support vector machine (SVM) classifier to classify the logos based on max pooling of local descriptors. Experiments show that the proposed method attains high recognition accuracies in decent time on logo images captured by surveillance cameras in the real-world scenario, which verifies the robustness and effectiveness of the proposed framework.