The Research Of Model Recognition Method For Multi-pose Vehicle

With the increasing number of motor vehicles in China,the traffic management department is facing great pressure and challenge.Accordingly,the identification of vehicle models is becoming a very important task.Large number of motor vehicles,small difference between some vehicle models,and different attitude appearing in pictures captured from different angels,all bring difficulty to the identification of vehicle models.In order to solve the multi-attitude vehicle model recognition problem,deep-learning based methods are adopted,and two fine-grained vehicle model recognition methods are proposed,one is based on FR-ResNet model and another is based on visual attention model.Experiments carried out on open multi-pose vehicle datasets StanfordCars and CompCars show that our methods can both improve the recognition effort of vehicle models.The main work of this paper is as follows:(1)Introduce the classic models of deep learning convolutional neural network,especially the residual neural network ResNet,and improve the structure of the classical residual network;introduce the fine classification data set of large vehicles commonly used at present.(2)A vehicle model fine identification network model FR-ResNet is proposed.Based on the improved residual network structure,the strategies of multi-scale input,reuse of low-level features in high-level layers and feature map weight learning are adopted to realize the reuse of features in improved network.FR-ResNet is a deep convolutional neural network model based on the reuse of residual network features,which achieves high recognition accuracy on both datasets.(3)A multi-attitude vehicle model recognition algorithm CAM based on visual attention model is proposed.CAM uses a convolutional neural network as the main body,which is combined with a local bilinear weighting structure.Based on the attention mechanism,LSTM is used to re-encode the global convolution feature map group;convolutional LSTM structure is embedded in the upper part of the network to learn high-level feature space relationships.CAM can be added to a variety of classical convolutional neural network models,and the comparison between the original network model and our improved CAM+ model can bring a lot improvement in recognition accuracy.