Abstract: Vehicle color is one of the significant vehicle details,and the recognition for it can provide more precise and rich information for vehicle identification in the Intelligent Transportation System.In the natural traffic scenes,the vehicle images obtained by cameras are greatly affected by the illumination changes.Therefore,the vehicle color can't be determined directly by the RGB values of the image.The traditional machine learning methods for vehicle color recognition require feature selection steps depending on experience,which may lead to limited classification effect.When applied to actual applications,these approaches probably have high computation cost and are difficult to obtain real-time results.Aiming at the problem that vehicle color information is difficult to gain and describe in natural scenes,a novel deep neural network model based on multi-color spaces is proposed to identify vehicle color in natural traffic scenes(MultiColor-Net).In the MultiColor-Net model,several filters of different sizes are used to extract the features of the input images both in RGB color space and in HSV color space,respectively.Then,the above features obtained in two different color spaces are combined together to get the classification results of vehicle color through a fully connected network.By comparing the experimental results of ResNet,Inception v3 and other deep neural network models with the MultiColor-Net proposed on real intelligent transportation data sets,the accuracy rate of MultiColor-Net is improved by about 2.45% with the HSV images alone,and the accuracy rate is improved by about 0.8% with the RGB images.Consequently,the proposed model,MultiColor-Net,can achieve a high recognition accuracy rate with the real traffic image data,and maintain lower computational complexity.

Key words: Color recognition, Deep neural networks, HSV color space, Intelligent transportation, Vehicle color