Vehicular Color Recognition with Dynamic and Complex Environment Based on WRCNN

Vehicular color recognition is a challenging problem since the change of light conditions, the interference of background color and difference subjective perceptions in dynamic complex environments. This paper proposes a method closing to the subjective cognitive model of human vision systems for vehicle color recognition. Firstly, a Spatial-Color (SC) normalized method is designed to extract the main color of vehicle images. Secondly, a novel Wide Residual Convolution Neural Network (WRCNN) is proposed to extract the global features, and the output is provided by a fully connected layer. Finally, A softmax classifier is used. Compared with those traditional color classification methods in which the color-spatial objective distance is calculated, the proposed method is more effective. Compared with AlexNet and VGG, our method decreases error rate by using deeper networks and residual structures, it also optimizes the constringency of networks. Experimental results show that, the training accuracy rate can reach 99.12% with 30, 000 training and 3, 600 testing images. The proposed method satisfies with practical real-time applications.