Image feature selection embedded distribution differences between classes for convolutional neural network

Abstract

Convolutional neural networks have achieved a great success in feature extraction and classification of images. However, some of the features extracted by convolutional neural networks are with insignificant difference between classes, which not only contribute little to image classification, but also increase the complexity of the classifier. It is important to select features that are helpful for image classification when using convolutional neural network. In view of the existence of class labels of image samples when training classifier, and motivated by the intention that these labels may also play a certain role in feature selection for image classification, we propose a feature selection approach by taking the distribution differences between classes into consideration on the basis of the features extracted by convolutional neural network. To be specific, we use the Gaussian mixture model to approximate the distribution of each feature on each subclass, and select the features significantly contribute to classification by designing a measure of distribution difference according to the numerical characteristics described by Gaussian mixture models. Further, an image classifier can be presented by redesigning the fully connected layers of the convolutional neural network based on the selected features. The proposed feature selection is adopted to image classification, and the experimental results show the effectiveness of the method.