Semisupervised PCA Convolutional Network for Vehicle Type Classification

In the vehicle type classification area, the necessity to improve classification performance across traffic surveillance cameras has garnered attention in research especially on high level feature extraction and classification. The backpropagation (BP) training approach of traditional deep Convolutional Neural Network (CNN) approach is time-consuming without using a Graphics Processing Unit (GPU). In this paper, we propose a semisupervised strategy for the end-to-end Principal Component Analysis Convolutional Network (PCN) in the area of vehicle type classification. Even without using a GPU, the proposed model eliminates the time-consuming training procedure of convolutional filter bank. In particular, the convolutional filters of the network are generated using unsupervised learning by Principal Component Analysis (PCA) which has tremendously reduced training cost and also reinforced the robustness of extracted features against various distortions. In order to further improve the training procedure while still preserving the discriminative characteristic of the system, only the fully-connected layer is fine-tuned in the supervised classification stage. The PCN is tested using a public BIT-Vehicle dataset which comprises 9850 surveillance-nature vehicle frontal-view images. The PCN can be easily implemented and readily compatible with many effective classifiers. Two classifiers, namely softmax classifier and Support Vector Machine (SVM) are employed in this network and their classification performances are then compared. Both classifiers take less than 100 seconds in the training process and are able to produce an average accuracy of above 88.35%, even under various inferior imaging conditions.