基于合作博弈Shapley值法的类激活映射算法  

Class activation mapping algorithm based on cooperative game Shapley value

作　　者：许莉 常雨晴 柴霁轩 宛旭 范纯龙 XU Li;CHANG Yu-qing;CHAI Ji-xuan;WAN Xu;FAN Chun-long(School of Computer Science,Shenyang Aerospace University,Shenyang 110136,China;Department of UAVs,Shenyang Aircraft Design and Research Institute,Shenyang 110034,China)

机构地区：[1]沈阳航空航天大学计算机学院,辽宁沈阳110136 [2]沈阳飞机设计研究所无人机部,辽宁沈阳110034

出　　处：《计算机工程与设计》2025年第3期795-803,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62171295)。

摘　　要：为加深对深度神经网络内部决策依据的理解,更好进行网络的调试和应用,提出一种结合特征重要性算法和类激活映射(CAM)的计算机视觉可解释性技术(Shapley-CAM)。利用合作博弈理论中的沙普利值算法计算特征图对最终结果的贡献,以此作为权重对特征图进行加权求和,得到类激活图,对神经网络模型的决策机制进行解释。重点考虑网络最后一层中每个特征图对结果的影响,可视化输入图像中对模型输出造成正向影响的区域。实验结果表明,该方法能够更准确地解释深度神经网络的决策依据,在定位能力和算法忠诚度等方面的性能得到了显著提升。To deepen the understanding of the internal decision-making mechanisms of deep neural networks for enhanced network debugging and application,a computer vision interpretability technique,termed Shapley-CAM,was proposed.Feature importance algorithms were integrated with class activation mapping(CAM).The Shapley value algorithm from cooperative game theory was leveraged to compute the contribution of feature maps to the final result.These contributions were used as weights for a weighted summation of the feature maps,resulting in a class activation map that explained the decision mechanism of the neural network model.Focusing on evaluating the impact of each feature map in the last layer of the network on the final outcome,regions in the input image that positively influenced the model's output were visualized.Experimental results demonstrate that this approach provides a more accurate interpretation of the decision rationale of deep neural networks,leading to significant improvements in performance metrics such as localization accuracy and algorithm fidelity.

关 键 词：深度神经网络 特征重要性 类激活映射 可解释性 合作博弈 沙普利值 特征图 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]