Vehicle Color Recognition Based On BP Neural Network

The aim of vehicle color recognition system is to recognize the color of vehicle accurately in real time. It can overcome many defects of traditional vehicle management system such as low efficiency and errors caused by man-made subjective factors, so that the automation level of vehicle management system can be greatly improved, which is of great significance in attacking vehicle crimes and mastering the situation of road traffic precisely. But the color of vehicle is often influenced by noises, illumination and so on, which affect accuracy of color recognition seriously. The research on the vehicle color recognition system is performed under this background. The main content and innovation of this thesis can be summarized as follows:Impulse noises in a color image are detected by the scalar method. Based on the result of impulse noises detection, an algorithm of improved median vector filtering is applied to remove the impulse noises. Removal of impulse noises is important for the following image processing. Locating the recognition area is the precondition of extraction for color characteristic information. By full utilizing of the correlation of vehicle-head and recognition area as well as axis symmetry of vehicle-head, the recognition area can be located through OTSU algorithm and twice projection.The color of highlight reflected from surface of vehicle seriously affects vehicle's true color. In order to improve accuracy of color recognition, the highlight pixels on surface of vehicle need to be detected before recognizing vehicle's color. In this paper, a new conception, named best diffuse reflection pixel, is put forward by using the input image and corresponding specular-free image. To detect highlight pixels within a single color real image, an algorithm is proposed according to the best diffuse reflection pixel. This algorithm is simple and easy in implementation. And its theoretical basis is dichromatic reflection model. Experimental results have shown that this algorithm is efficient and fast, which can meet requirements of accuracy and real-time processing in the vehicle color recognition system. The accuracy of vehicle color recognition is improved effectively under strong sunshine. A BP neural network is designed to recognize the color of vehicle. Traditional BP algorithm has many defects, such as slow rate of convergence, falling into a local minimum easily and so on. Network parameters are needed adjusted to avoid these defects. The color of vehicle is transformed from RGB color space to I1I2I3 color space which has better effect in pattern recognition. The feature vectors are then standardized to art as the input vectors of BP neural network. Trained samples are preprocessed by fuzziness at first. The expected output vectors are evaluated by the fuzzy membership value. Experimental results have shown that trained network can recognize vehicle color effectively, which conforms to human visual characteristic.