影像解译中的深度学习可解释性分析方法

深度学习的迅速发展极大地推进了多种图像解译任务的精度提升,然而深度学习网络模型的“黑箱”性质让使用者难以理解其决策机理,这不仅不利于模型结构优化和安全增强等,还会极大地增加训练调参成本。对此,本文围绕影像智能解译任务,对深度学习可解释性国内外研究进展进行了综合评述与对比分析。首先,将当前可解释性分析方法分为激活值最大化分析法、代理模型分析方法、归因分析法、扰动分析法、类激活图分析法及样例分析法等6类方法,并对方法的原理、侧重点及优缺点进行了回顾。其次,对8种衡量各类分析方法所提供解释的可靠性的评估指标进行了回顾,并梳理了当前公开可用的可解释性开源算法库。在当前开源算法库的基础上,以遥感影像智能解译任务中的可解释性分析为例,验证了当前深度学习可解释性方法对遥感影像的适用性;试验结果表明当前可解释性方法在遥感解译中还存在一定的局限性。最后,总结了现有基于自然影像的可解释性算法在遥感影像解译分析中存在的问题,并展望了设计面向遥感影像特性的可解释性分析方法的发展前景,旨在为相关研究者提供参考,推动面向遥感影像解译的可解释性方法研究,从而为深度学习技术在遥感影像解译任务中的应用提供可靠的理论支持与算法设计指导。

Deep learning interpretability analysis methods in image interpretation

The rapid development of deep learning has greatly improved the performance of various computer vision tasks. However, the black box nature of deep learning network models makes it difficult for users to understand its decision-making mechanism, which is not conductive to model structure optimization and security enhancement and also greatly increases the training cost. Focusing on the task of intelligent image interpretation, this paper makes a comprehensive review and comparison of the research progress of deep learning interpretability. Firstly, we group the current interpretability analysis methods into six categories: activation maximization method, surrogate model, attribution method, perturbation-based method, class activation map based method and example-based method, and review the principle, focus, advantages, and disadvantages of existing related works. Secondly, we introduce eight evaluation metrics that measure the reliability of the explanations provided by the various interpretability analysis methods, and sort out the current publicly available open source libraries for deep learning interpretability analysis. Based on the open source library, we verify the applicability of the current deep learning interpretability analysis methods to the interpretation of remote sensing images. The experimental results show that the current interpretability methods are applicable to the analysis of remote sensing interpretation, but have certain limitations. Finally, we summarize the open challenges of using existing interpretability algorithms for remote sensing data analysis, and look forward to the prospect of designing interpretability analysis methods oriented to remote sensing images. We hope this review can promote the research on interpretability methods for remote sensing image interpretation, so as to provide reliable theoretical support and algorithm design guidance for the application of deep learning technology in remote sensing image interpretation tasks.