Vehicle Image Recognition Algorithm Based On Deep Transfer Learning

In the context of big data,the lack of complete data for annotations makes traditional deep learning image recognition methods have some problems,such as too many steps,high threshold,and poor generalization.As a branch of machine learning,transfer learning tends to transfer the knowledge trained on big data to the proposed new recognition task.Using a small number of new samples and iterations can achieve the desired effect,which greatly reduces the training threshold of deep learning.At present,most of the road monitoring and various vehicle identification instruments are reserved for small cars and license plate recognition,lacking a wider range of application scenarios.Moreover,in the face of the market demand for continuous upgrading of vehicle types,and the specific road conditions of different roads are quite different.Under the premise,it is always necessary to redesign the deep learning model from scratch after customizing it to the relevant enterprise.Therefore,it is necessary to design a model that can easily increase or decrease sample categories according to personalize need.Using only a small number of samples and a small amount of training time,the model can achieve good results with good generalization and portability.Firstly,select a feature extraction network with robust portability.Take several convolutional network models as the backbone network.Import the pre training parameters on Image Net on the basis of maintaining the original structure of the network.Quickly extract features by freezing part of the network layer.Design a new classifier,and carry out domain adaptation on the VOC2012 data set that selects the required categories.In the experimental results,the recognition accuracy of each backbone network after applying transfer learning has increased by 0.9-14.4% compared to the performance in Image Net.The training time on CPU after freezing the feature layer is also reduced by more than 50%.Select the optimal backbone network,and then,compare the effects of single and secondary transfer on the vehicle data set.The experimental results show that the source domain with higher domain similarity can further improve the training efficiency of secondary transfer on the basis of the former.And the recognition accuracy is improved by about 11.8-14% compared to the performance in Image Net.In order to further improve the generalization of the model and speed up the training process.Based on the feature extraction network Efficient Net V2 and corresponding parameters obtained in the above step,the improved subspace transformation method CORAL(Correspondence relation Alignment)is applied to different depths of the model to make the model fit in new tasks with different feature distributions and speed up convergence.Finally,the model is loaded with the weights of the three source domains and compared with other representative transfer learning methods on the Office31 dataset.The results show that the deep transfer model in this paper is better than other transfer learning methods compared,and there is no overfitting phenomenon under the condition of as few samples as possible.The training efficiency is much better than ab initial training.CORAL can not only converge quickly,but also improve the final recognition ability of the model.