FastScore-CAM:一种高效的视觉解释方法  

FastScore-CAM:An Efficient Saliency Visual Interpretation Method

作　　者：袁月璨 郑熠 张烺 YUAN Yuecan;ZHENG Yi;ZHANG Lang(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《软件导刊》2025年第4期108-114,共7页Software Guide

基　　金：四川省科技计划资助项目(2024ZDZX0007)。

摘　　要：在深度学习模型应用中,可解释性对于建立用户对模型的信任至关重要。视觉解释方法Score-CAM利用特征图扰动输入样本,并与特征重要性权重进行线性组合,表现出较高的准确率和跨模型的稳定性。然而,Score-CAM中的扰动样本需要反复调用模型而导致算法效率降低、响应缓慢。为了解决该问题,提出一种高效的基于扰动的视觉解释方法——FastScore-CAM,该算法在保持Score-CAM算法性能的同时可缓解运行效率低下的问题。进一步通过研究分析Score-CAM的基本工作原理,引入了非负矩阵分解(NMF)针对性地减少扰动样本并过滤冗余特征。实验结果表明,该方法不仅保持了较高的准确率和跨模型的稳定性能,还显著提高了原方法的运算速度。In the application of deep learning models,interpretability is crucial for establishing users'trust in the model.The visual explana‐tion method Score-CAM perturbs the input samples using feature maps and linearly combines them with feature importance weights,demon‐strating high accuracy and cross-model stability.However,the perturbed samples in Score-CAM require repeated calls to the model,resulting in reduced algorithm efficiency and slow response.To address this issue,an efficient perturbation-based visual explanation method,Fast‐Score-CAM,is proposed.This method alleviates the problem of low operational efficiency while maintaining the performance of the Score-CAM.Specifically,by studying and analyzing the basic working principle of the Score-CAM,Nonnegative Matrix Factorization(NMF)is in‐troduced to specifically reduce perturbed samples and filter redundant features.Experimental results show that this method not only maintains high accuracy and cross-model stable performance but also significantly improves the computational speed of the original method.

关 键 词：Score-CAM 可解释性 深度学习 视觉解释方法 显著性图 非负矩阵分解 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]