基于GAN的图像识别系统蜕变测试方法

图像识别是图像处理的重点研究领域，在测试结果难以判定以及数据集样本类别不平衡的影响下，适用于图像识别系统健壮性以及稳定性的测试技术较为欠缺。为有效测试图像识别系统，本文提出将蜕变测试方法用于图像识别系统的测试过程中，依据生成式对抗网络来生成贴近现实的衍生数据以构建适用于图像识别系统的蜕变关系，引入衍生图像质量验证方法与测试结果自动判断方法以构建面向图像识别系统的蜕变测试框架流程。最后，通过车辆识别案例研究验证本文方法的可行性与有效性。实验结果表明，本文方法能检测出图像识别系统在不同场景中的不一致行为，能有效地评估系统健壮性。

Method of Metamorphic Testing for Image Recognition System Based on GAN

Image recognition is a key research field of image processing. Under the influence of the difficult determination of test results and the unbalanced category of data set samples, the test technology suitable for the robustness and stability of image recognition systems is deficient. In order to effectively test the image recognition system, this paper applies metamorphosis test method into testing process of image recognition system, generates derived data close to reality based on Generative Adversarial Network to construct a metamorphic relationship suitable for the image recognition system, introduces derivative image quality verification methods and automatic judgment methods of test results to construc a metamorphic testing framework process for image recognition systems. Finally, the feasibility and effectiveness of this method are verified through a case study of vehicle identification. The experimental results show that the method can detect the inconsistent behavior of the image recognition system in different scenes and can evaluate the robustness of system effectively.