Fine-grained Vehicle Model Recognition Based On Deep Learnin

Fine-grained vehicle type recognition plays an important role in intelligent transportation.Its purpose is to recognize the manufacturer,model,year and other information of the vehicle image.The accuracy of vehicle type recognition is not high due to the large appearance difference caused by the same model due to the different attitude and perspective,the small appearance difference between different models and the small number of fine-grained vehicle data sets.Therefore,this paper aims to improve the accuracy of fine-grained vehicle type recognition by designing convolutional neural network(CNN)with strong feature extraction capability and using data augmentation technology to expand the number of experimental samples.The main research contents are as follows:(1)In order to enhance the feature extraction capability of CNN,this paper proposes a vehicle type recognition network based on the channel space confounding attention mechanism.There are many methods based on attention mechanism.In order to select the best attention mechanism,CNN firstly introduces three improved attention mechanism modules: channel attention module,spatial attention module and channel spatial obfuscated attention module,so that CNN can pay attention to local features of vehicle images.Secondly,the embedding mode of attention module has a great influence on model performance,so this paper proposes three embedding modes of attention module: serial mode,parallel mode and residual mode.Through comparative experimental analysis,it is concluded that the embedding of channel space confounding attention mechanism module in parallel mode can make CNN pay attention to more discriminative local information and bring the optimal promotion effect for CNN.(2)In order to further improve the accuracy of fine-grained vehicle type recognition and solve the problem of poor enhancement effect of random data,this paper proposes a data enhancement network based on semantic information.The network constructs a local attention module through bilinear attention gathering to obtain representation data with semantic information.Because this kind of data pays attention to the local details of the vehicle,it can effectively increase the training data of the network and improve the accuracy of the model.In addition,in order to enhance the feature expression ability of the model,the channel space obfuscation attention module is embedded in parallel before the local attention module.Experiments show that the semantic data enhancement method proposed in this paper can further improve the accuracy of fine-grained vehicle recognition.(3)According to the experiment,it is found that the discriminant force is not considered in the confused attention of channel space,so this paper proposes a vehicle type recognition network based on counterfactual attention.The network uses counterfactual causality to learn attention and provides a powerful tool for evaluating the quality of attention,thus guiding the network to learn more discriminative fine-grained features.In addition,in order to improve the forward propagation capability of the model,Meta-ACON activation function is used to adaptively select neurons to be activated.Experiments show that the counterfactual attention network proposed in this paper has the best effect on fine-grained vehicle recognition.