In the interpretability research of convolutional neural network (CNN), the quantitative analysis of feature information is the focus of the research. In this paper, we proposed a feature measurement method based on information entropy to quantitatively analyze the performance of feature extraction in CNN. Firstly, aiming to the different types of images, the activation histogram of the feature layer is calculated, and then the normalized entropy is calculated to define the characteristic purity. Different CNN models and their feature purity of internal structures were quantitatively evaluated on CIFAR10 and ImageNet datasets. At the same time, visual interpretation was performed by combining class activation maps to verify the relationship between feature extraction ability and feature purity of a specific CNN model. The experimental results show that the feature purity is significantly correlated with the feature extraction ability and classification performance of the model. At the same time, the calculation of feature purity is not dependent on the classification label, is not limited to the specific network structure, and has strong robustness and practical value. The proposed quantitative evaluation method can effectively evaluate the feature extraction performance of CNN.