一种基于蜕变测试的卫星遥感目标检测模型鲁棒性测试方法

A robustness testing method based on metamorphosis testing for satellite remote sensing object detection models

通过深度神经网络等智能方法实现对卫星遥感影像的目标定位识别成为当前研究与应用的主流。针对其广泛落地应用对软件测试工作带来的诸多挑战,围绕鲁棒性这一重要的测试属性,提出一种基于蜕变测试的卫星遥感目标检测模型鲁棒性测试方法。该方法构建“淡云层遮挡不改变模型输出结果”“明暗度变化不改变模型输出结果”两种蜕变关系,通过传统CV方法生成衍生用例,并提出量化指标进行评估。在TGRS-HRRSD数据集两类小目标上分别评估基准模型和改进模型,结果表明,该方法在评估模型鲁棒性上,具有一定的有效性与直观性。

Target recognition and localization of satellite remote sensing images through deep neural networks and other intelligent methods has become the mainstream of current research and application.In view of the many challenges to software testing brought about by its widespread application,a robustness testing method based on metamorphosis testing for satellite remote sensing object detection model is proposed,centering on the important test property of robustness.The method constructs two kinds of metamorphic relationships,namely,“light cloud cover does not change the model output result”and“brightness change does not change the model output result”,generates derived use cases by traditional CV method,and proposes quantitative indexes for evaluation.The baseline model and the improved model are evaluated on the TGRS-HRRSD dataset with two types of small-objectives,and the results show that the method is effective and intuitive in evaluating the robustness of the model.