图片分类的卷积神经网络可解释性分析

Interpretability Analysis of Convolutional Neural Networks Based on Image Classification

为了理解卷积神经网络在图片分类任务中做出决策的依据,进而优化模型,降低调参成本,对卷积神经网络进行可解释性分析是十分有必要的。为此,文章以水果图片分类任务为切入点,使用了多种类激活图,从多个角度分析模型所给出结果的原因。文章采用Res Net模型先进行微调,在取得较好的分类性能后,进行了语义特征的基础分析、遮挡性分析,以及基于CAM的可解释性分析和LIME可解释性分析,为卷积神经网络提供一定的可解释性。实验结果表明,卷积神经网络做出决策的依据与人类理解的语义是一致的。

In order to understand the basis for decision making of convolutional neural network in image classification task,so as to optimize the model and reduce the cost of parameter adjustment,it is necessary to analyze the interpretability of convolutional neural network.For this reason,this paper takes the fruit image classification task as the starting point,uses multiple kinds of activation graphs,and analyzes the reasons for the results given by the model from multiple perspectives.In this paper,ResNet model is used to fine tune and achieve better classification performance.The basic analysis of semantic features,occlusion analysis,CAM based interpretability analysis and LIME interpretability analysis are carried out to provide a certain interpretability for convolutional neural networks.The experimental results show that the decision basis of convolutional neural network is more consistent with the semantic concepts understood by human beings.