| In recent years,with the significant improvement of computer capacity,the technology of deep neural network has achieved rapid development,and has made a qualitative leap in image recognition.However,the deep learning model needs enough training samples in the training process to be able to better generalize different samples,and the training time is long and the training efficiency is not high.At the same time,if you want to obtain some image data,it is also a troublesome thing,and the cost of annotation is relatively large,which leads to the difficulty of obtaining some data.Based on this,people expect to solve the small sample problem.Small sample learning is to identify objects through a small number of samples,which can be used to collect data for tasks requiring a large amount of training data,and small sample learning reduces the difficulty of obtaining data sets.This paper studies the problems in small sample image recognition,and introduces how to improve the generalization ability of the model to improve the accuracy of model image recognition.The content of this paper is as follows:(1)Meta learning based on deep learning can learn known categories,and metric learning based on meta learning can embed features,map feature information into the feature space,and then identify new test tasks.However,many current measurement models have poor generalization ability,and the network model has insufficient feature expression ability.Based on this,this paper improves the structure of the small sample classical relational network,and adds the inception block to the embedded module in the relational network,thus enhancing the feature extraction ability of the embedded module.The experiment on the Mini Image Net dataset commonly used for small samples verifies that the inception block can enhance the feature extraction ability of the network,so as to improve the accuracy of model recognition.(2)The relational network in small sample learning is easy to cause gradient explosion when the gradient drops,and the gradient flow is unstable.Therefore,in order to stabilize the gradient flow,make the model avoid gradient disappearance and gradient explosion,and make it more efficient when backpropagation and gradient descent occur,this paper improves the activation function and loss function in the relationship network.Changing the Relu activation function into the Mish activation function in the original relational network will retain a smaller negative value as much as possible,so that the gradient flow of the network can be stabilized,and better information flow can be achieved.The mean square error loss function in the original relational network is replaced by the mean absolute error loss function.In the training process,there is a stable gradient,and there will be no gradient explosion,which is convenient for better training of the model.(3)Through experiments on three commonly used small sample learning datasets of Omniglot,Mini Imagenet and Tiered Image Net,the results show that the improved network has higher recognition accuracy than the original relational network model.The effect on the Omniglot dataset has been enhanced.The accuracy rate on the 1-shot task of the Mini Imagenet dataset has been improved by 3.8%,the 5-shot task has been improved by 3.73%,the 1-shot task of the Tiered Image Net dataset has been improved by 4.21%,and the 5-shot task has been improved by 4.05%.This confirms that the improved network improves the accuracy of small sample image recognition,and can effectively enhance the generalization ability of the model. |
