Dimension-reduction fine-tuning method for image classification of few samples

Abstract

The image classification problem of few samples is studied. The fine-tuning image classification method of deep convolutional neural network is analyzed, and it is pointed out that the image feature vector extracted by deep convolutional neural network can be projected into low-dimensional space, which increases the sample density. Based on this, a dimension-reduction fine-tuning method for image classification problem of few samples is proposed, which uses the principal component analysis algorithm to reduce the dimension of image features extracted by the pre-trained convolutional neural network, and then trains the classifier with the dimension-reduced feature vectors. This method improves the classification accuracy when the number of samples is very small, and is analyzed and verified by simulation.