The Research on Classification of Small Sample Data Set Image Based on Convolutional Neural Network

This paper is aimed to study the problem of small sample image recognition. We use the initialization method of dictionary filter to replace the traditional Convolutional Neural Network (CNN) initial filter method so that the dictionary filter can learn its model structure during the CNN training process and the CNN model can extract a more Multi-effective feature map. For the image classification stage, using GRNN to replace BP neural network can enhance the ability of model classification processing, accelerating the speed of model convergence, and comprehensively improving the recognition accuracy of the model. This article extracted a small amount of data in large data set for experiments, and use the ORL face database as verification simultaneously, the experimental results and analysis are given. Compared with the fine-tuning through migration learning, it has increased by 3.84%, compared with the use of fully connected classifiers, it has increased by 4.58%, which proves the accuracy of the proposed method.