Fine-grained Recognition Of Car Models In Low-resolution And Multi-angle Monitoring Scenarios

With the rapid development of computer vision technology,convolution neural network is used in the field of vehicle recognition,and high recognition accuracy is achieved.However,the current research focuses on the recognition of high-definition vehicle image,and the research on the recognition of low resolution vehicle image in the monitoring scene is less.Due to the low resolution of the image collected from the monitoring scene,the image contains less vehicle feature information,which increases the difficulty of fine-grained vehicle recognition.Therefore,the scene of fine-grained vehicle recognition has a large research space,can bring some practical value.In view of the above problems,this paper introduces the image super-resolution reconstruction algorithm.In order to achieve high real-time performance,the network structure and loss function are improved.At the same time,the original low resolution data set is reconstructed.A bilinear vehicle recognition network based on multi-scale feature fusion is proposed,which improves the accuracy of low resolution vehicle image recognition in the monitoring scene.The main contents of this paper are as follows:First of all,due to the redundancy problem in the network generation of image super-resolution algorithm against neural network,the fire module network structure is adopted,and the compression layer and expansion layer are redesigned,which effectively reduces the network redundancy problem,and reduces about 40% of the parameters compared with the original network.The original network normalization layer loses the corresponding image details and affects the final image quality,so the normalization layer is removed.In this paper,an Otsu adaptive thresholding segmentation method is proposed.The PSNR value of the binary image of the original image and the binary image of the generated image is calculated as a loss function of the generated network.Then,the SR-Box Cars116 k dataset is obtained by super-resolution reconstruction of Box Cars116 k,which is a low resolution multi angle surveillance scene dataset.In order to solve the problem that the number of dataset images is too small,occlusion,color change,inversion,local interception and other methods are used to enhance the dataset,which expands the number of dataset images to a certain extent.Using unpack method for vehicle image preprocessing can effectively reduce the background noise and increase the proportion of the target image in the overall image.Finally,a bilinear feature extraction network is designed to extract the features of vehicle image.Res Net is used as network A for feature extraction and VGG19 as network B.At the same time,the idea of multi-scale feature fusion is adopted,and the multi-scale fusion method is used in network B to fuse the low-level semantic feature information with the high-level semantic feature information.Because the dimension and scale of each layer are inconsistent,convolution layer and pooling layer are designed to fuse all features.The global average pooling layer and dropout layer are used to reduce some parameters.By using focal loss function,the imbalance of sample distribution is solved to a certain extent.At last,the recognition rate of this network is better than that of the classical network.