Vehicle image recognition algorithm based on EfficientNetV2 model transfer learning

In intelligent transportation system, due to the problems of large vehicle mobility and poor road shooting environment, most of the traditional deep learning can only recognize the number of vehicles and license plates, and the data set in this field is also extremely scarce. Based on these problems, an image recognition model with EfficientNetV2 as the backbone network is proposed. Using the pre training parameters of ImageNet, and freezing part of the network to quickly extract image features. Select the transportation facility related parts in VOC2012 data set for domain adaptation training, and get a vehicle recognition model with strong portability. Take the model parameters obtained in the previous step as the source domain, and transfer the parameters to the vehicle data set again. Compared with the single transfer of ImageNet as the source domain, it can be seen that the domain with higher domain similarity has better transfer effect. Finally, based on the feature extraction network and corresponding parameters obtained in the previous step, combined with the subspace transformation method Coral, the features of different network depths are constrained, so that the model can be adapted to new tasks with different feature distributions and accelerate convergence. A small number of samples are used to detect whether the model is overfitting. Through experiments, it can be seen that the recognition accuracy and training speed of the model are greatly improved after transfer learning is used in a small sample data set, and it can be easily reused for recognition tasks in other similar fields.