Research And Implementation Of Fine-grained Car Model Recognition Based On Convolution Neural Network

Fine-grained car model recognition is of great application value in intelligent traffic management,illegal vehicle detection,automobile brand market statistics and marketing and other fields.However,due to the complexity of fine-grained classification problem,the accuracy of fine-grained car model recognition still has much room for improvement.Based on the current popular convolution neural network and its variants,the problem of fine-grained car model recognition is studied in this paper.On the one hand,the accuracy of fine-grained car model classification is improved through the research and improvement of algorithms;On the other hand,this paper designs and implements a function-rich,easy-to-use fine-grained car model recognition system test software.The main work of this article is summarized as follows:1.Through the research and comparison of the object detection algorithm,the SSD algorithm is selected to detect the vehicle area in the picture,thereby eliminating the background interference.Taking the fine-grained car model dataset Model_dataset collected in this article as the experimental data,experiments were conducted using two classic networks: CaffeNet and Vggnet-16.The experimental results show that for the two neural networks,the fine-grained car model classification accuracy after removing the background interference is increased by 10.49% and 8.91% compared with using the original images.In the process of researching SSD algorithm,this paper proposes a method to modify the SSD network structure by analyzing the data distribution and convolution output conditions for specific dataset.The experimental results show that both converge speed and detection speed of the object detection algorithm model after the modification have improved.2.From the perspective of transfer learning,this paper selects the car type classification as the source task for the fine-grained car model classification of the target task,and collects a larger-scale source task dataset Type_dataset through the web crawler.By migrating the “knowledge” obtained from the source task to the target task,the purpose of improving the classification accuracy of fine-grained car models is achieved.The simulation experiment results show that compared using the ImageNet pre-training models,the accuracys of the CaffeNet network and Vggnet-16 network fine-tuned on the testset are improved by 2.28% and 0.27% respectively on the basis of the source task pre-training models designed in this paper.3.A quantified car pose feature representation scoring model is established.Based on the criteria,different network models and convolutional features of different levels are evaluated to select the best car pose feature representation.Using the selected pose features,the k-means clustering algorithm is used to cluster the car images,and the car pose is divided into 8 kinds of attitude according to the head orientation,which are "left","left front","front","right","right","right rear","after" and "left behind".Experiments show that the accuracy of car pose recognition using the above method is as high as 91.41%.At the same time,it is transplanted to Hadoop distributed platform for the characteristics of large space occupied by convolutional features and low computational efficiency of k-means algorithm.Through simulation,it is proved that this method can effectively reduce the stand-alone calculation pressure and disk pressure in the process of car pose clustering.4.Using the HTML,JavaScript,CSS front-end language and the Python back-end framework Tornado,a fine-grained model car recognition system testing software is designed and implemented,which combines the functions of car detection,pose division,car model classification,and car type classification.