Convolutional Neural Network have considerable interpretability in the image recognition process,enabling the extraction of relevant feature concepts in an understandable way.This paper explores an approach to exploit the interpretability of Convolutional Neural Network,and proposes a semantic fea-ture-oriented distance loss function.It forces the base models to learn different target feature concepts and the model fusion approach integrates the relevant semantic features to recognize the target object.The cur-rent Cifar10 data set and VGG16 network are used as benchmarks,the method framework is explored from key components such as initialization,distance function,segmentation threshold,fusion method and image enhancement.As a result,the influence mechanism of relevant components on the method framework is clarified,and the good improvement effect of the method on image recognition ability is confirmed.
维普
万方数据