Deep Learning Model Construction Based On Small Samples

Few-sample learning is one of the research methods to deal with computer vision problems,and the transfer learning model based on the idea of small-sample learning can complete computer vision target tasks with a small number of samples.Since transfer learning models perform well in image classification tasks with small sample sizes and similar data scales but different target domains,they have been widely studied and applied to real-world problems.However,due to the poor generalization ability of the transfer model for new classes in small sample data,and the insufficient ability to extract and transfer deep features,there are still many improvements and improvements in this research direction.Aiming at the problem that the target domain and source domain in transfer learning are small samples,which leads to the low accuracy of deep feature transfer recognition in the deep learning model,and the model is easy to overfit.Based on the deep pre-adaptation network(DAN),this paper uses the genetic algorithm(GA)to optimize the network structure.A deep transfer learning model based on GA is proposed and compared with other transfer learning methods in several small-sample standard datasets.The work and main research results of this paper are as follows:(1)Aiming at the problems that deep model training is prone to overfitting under small samples,and the learning curve is unstable and jumpy,a deep migration Alex Net is proposed.First,the structure of the first five layers is frozen,and the last three layers are fine-tuned.,and make the training samples have similar input data and scale,and the target information is different from the cluster center,to prove the difference in the anti-overfitting ability and learning effect of this model compared with the non-transfer learning deep model.Experiments show that this model has stronger anti-overfitting ability and faster convergence speed.(2)In order to overcome the problem that the target domain and the source domain are small samples in the transfer learning,the deep feature transfer recognition accuracy in the deep learning model is low.A GA-DAN model is proposed,which uses the accordion coding paradigm to simplify the chromosome structure,improves the DAN structure and connectivity represented by the GA coding method,and conducts theoretical analysis and experiments on multiple standard small-sample migration datasets.The results show that the model can still effectively resist overfitting and has a faster convergence speed and higher accuracy under a small sample set.