Research On Fine-Grained Car Recognition Based On Deep Semantic Features Enhancement

With the rapid development of the motor vehicle market,the number of motor vehicles in our country has exploded.Traditional car type analysis and license plate number identification have failed to form a comprehensive and three-dimensional data structure to meet the increasingly complex road traffic requirements.Therefore,comprehensive and fine-grained car information collection problems need to be solved urgently.Fine-grained car recognition refers to the identification of manufacturers,brand models,and annual information through the car appearance images at any angle and scene,which is of great significance in the fields of intelligent traffic,security and many other fields.Fine-grained recognition tasks have subtler visual differences between different categories and are mostly concentrated in local areas compared with conventional classification tasks.Therefore,localization of local area and feature extraction are the keys to the performance of fine-grained recognition.Such subtle differences can hardly be characterized by traditional hand-crafted features.In this dissertation,based on convolutional neural networks,the fine-grained recognition method is studied in-depth.The recognition method of convolutional neural networks is enhanced by discriminable local features,which improves the recognition accuracy of fine-grained models and satisfies the requirements of car identification under real traffic scenarios.The main contents of this dissertation are as follows:1.The background and current research of convolutional neural networks and fine-grained recognition methods are introduced,and the convolutional neural network is analyzed and studied in-depth,including key technologies and important theories.The composition,characteristics and optimization methods of convolutional neural networks are explained in detail.Then,the principles and characteristics of several classical convolutional neural networks are demonstrated.The common methods for fine-grained recognition and its characteristics are analyzed systematically,and the difference between traditional methods based on hand-crafted feature extraction and deep learning methods are compared in detail.2.Traditional convolutional neural network models lack the representation of local semantic information,which limits the capability of the representation of image features.In this dissertation,we propose a convolutional neural network model based on semantic information fusion,combined with the methods based on object detection and feature fusion of deep learning to enhance the characterization of distinguishable semantic components and apply to fine-grained vehicle identification.The model consists of a localization network and a recognition network.The localization network obtains the specific location of the car object and the semantic components through the Faster RCNN.The recognition network firstly uses the convolutional neural network to extract the car's overall characteristics and local semantic part features.Then the features are spliced and integrated.Finally,the recognition result is obtained through the deep neural network.The use of small-kernel convolutions as a feature fusion of neural networks is proposed and the performance is verified on multiple public car data sets.The experimental results show that the method has achieved a good recognition accuracy.3.Fine-grained recognition methods based on regional localization and feature extraction are difficult to use in real-world scenarios due to the requirements of pre-given specific location of fine-grained objects and massive bounding boxes of local regions,and only applicable to a single object scene.A fine-grained recognition method based on region proposal network is proposed to solve these problems.The method first extracts the deep convolution features of the image by the convolutional neural network,and then generate the region candidates on the convolution feature map with the sliding windows.Next,the features of the region candidates are passed through classify layers and regression layers to obtain the object's probability and position respectively.Finally,the specific category and the exact location of the object is obtained by passing the candidate regions through the object detection network.The final recognition result is obtained by the non-maximum suppression algorithm.The experimental results show that the proposed algorithm overcomes the dependence of the traditional fine-grained recognition methods on the object location and achieves the fine-grained recognition of the car model in a complex scenario with better robustness and practicability.