| Vehicle automatic identification system has an important significance for traffic monitoring,parking management and public security system.The research for vehicle automatic recognition is based on computer vision and it is very deep and extensive.In these studies,vehicle detection and vehicle identification are important components of automatic identification systems.However,vehicle detection and identification has great challenges due to the lighting changing,different weather condition and the distortion of captured images in natural road scenes.Since 2012,with the significant achievements of deep learning in the field of image classification,convolution neural network has become a major topic in many field.In this paper,we studied convolution neural network in vehicle detection and vehicle recognition for real road scenes based on above studies the main work in this paper includes:1.This paper presents a vehicle detection algorithm based on feature map of convolution neural.By weighted feature maps in the convolution layer and weight parameter in the full-connection can accurately locate the vehicle location and recognize vehicle in the real road scene.The experimental results show that the average accuracy of the feature detection is very good and it has good robustness and applicability for images with noise interference such shade,fence.2.Based on AlexNet network,we studied the optimization problem in convolutional neural network,including the number of fully connected layers,the number of convolutional layers,the size of convolution kernel,the number of convolution kernels.And we also studied their impact on the classification accuracy of convolutional neural network,and their impact on the feature extraction time,network training time,0m model parameters.3.Vehicle classification network can only be classified according to vehicle make and can not be distinguished between the same brand of different color vehicles and different types of vehicles,so in this paper we presents an improved vehicle recognition network,by splitting fully connected layer in the network,and using multi-loss method to calculate the probability of different attributes.And we realized the recognition of vehicle brand,vehicle type and vehicle color for vehicle multi-attribute classification.4.This paper establishes a real scene image dataset,which contains 15 vehicle makes and a total of 15000 images. |
