Study On Vehicle Color Recognition Method With Convolutional Neural Networks In Complex Environments

Vehicle color is a key feature of the vehicle and also plays an important role in vehicle identification when comparing with the identifying of vehicle license plates,vehicle brand and vehicle logos.It can not only increase the reliability of vehicle ID recognition,but also help more in identifying illegally modified cars.What's more,it can provide certain function in the vehicle bayonet security systems inspection and cracking down on criminal activities for the security department.It is also an important key issue in the recognition of complex environment,such as illumination effects,dust and foggy weather.In the existing research on color recognition of vehicles,there are several problems.On the one hand,the recognition accurate rate of the vehicle color can be affected because the segmentation of color regions is not fine enough,such as image background,nonprimary color area of the car body and other factors,which affect the accuracy of positioning and segmentation.On the other hand,different vehicle manufacturers use different painting and coloring to make the color matching more and more complicated,the dust accumulation of the body over the years or the color change caused by oxidation of the coating and the color change under different illumination,etc.The degree of discrimination of the various color categories is increased,so that the classification recognition algorithm does not work well.Therefore,in this paper,for the research of vehicle color recognition based on convolution neural network in complex environment,we propose following two aspects:(1)Vehicle color region of interest segmentation.We proposed a regional saliency-based anti-jamming segmentation algorithm,which uses the multi-channel local-sensitive histogram transformation and multi-channel saliency detection to segment the vehicle image after the detection network is detected,and eliminates the interference region to obtain Vehicle body color area,which we call the vehicle color interest area.(2)A color classification method based on convolutional neural networks.We propose a convolution neural network model based on multi-color space.In three different color spaces,RGB,LAB and HSV,we construct a convolution neural network based on VGGNet and do some works in terms of data input processing and model output selection.We design a model output selection judger to fuse the prediction results of different color spaces and optimize the final prediction category of the output.The experimental results show that the color region segmentation algorithm we proposed can basically verify its effect and reach the state-of-the-art in some parts.Compared with the previous research,the convolutional neural network color classification algorithm pays more attention to the processing of sample data while performing well in the public and non-public data sets.