A PCANet based method for Vehicle Make Recognition

This paper proposed a new Vehicle Make Recognition (VMR) method using the PCANet features extracted from vehicle front view images. The PCANet architecture processes every input vehicle image through only three very simple data processing components: cascaded principle component analysis (PCA), binary hashing, and block-wise histograms, and generates a sparse vector as the feature representation. Then, a linear SVM classifier collects the PCANet features to train a model for the classification. For the evaluation of the proposed method, we built two large training datasets: a sedan dataset and a SUV dataset respectively. The sedan training dataset consists of 4188 sedan images corresponding to 22 different vehicle makes, and the SUV training dataset contains 2165 SUV images corresponding to 21 vehicle makes. Moreover, every make includes nearly all the models with different styles and colors in recent years. Compare this method with other two state-of-the-art VMR classification methods, we conclude that our approach outperforms other approaches and achieves high accuracies of 95.48% and 95.84% on our datasets, with a recognition speed of 0.4s per image.