可控图像退化对提升机器视觉鲁棒性的研究

现有工业机器视觉设备验证仍主要依赖批量造件,小概率的产品缺陷误判和设备失效风险不容易在投产前被发现和消除。为提升机器视觉系统鲁棒性,充分验证算法和参数不足,需要开发一种可控图像退化模拟系统。此系统能根据工况和设备特性,利用少量样图生成各类具有代表性的退化图像,并对图像退化程度进行精确量化评价。选择评价合适的退化图像对主流机器视觉设备进行鲁棒性验证,从而虚拟但有效地降低设备损坏停机和产品质量风险。另外,对于深度学习视觉应用,此系统也能更贴合工厂环境实现数据增广,优化神经网络模型。

Research on Controllable Image Degradation to Improve Machine Vision Robustness

Generally current industrial machine vision equipment verification still mainly rely on massive production, resulting in a small probability of misjudgment for product defects and equipment failure risks, which are not easy to be discovered and eliminated before being put into production. In order to improve the robustness of the machine vision system and fully verify the insufficiency of algorithms and parameters, it is necessary to develop a controllable image degradation simulation system. A variety of representative degraded images can be generated by the system originated from a small number of sample images according to the working conditions and equipment characteristics, while the degree of image degradation for these images is accurately quantified and evaluated. Appropriate degraded images are chosen to imply robust verification of mainstream machine vision equipment, to avoid equipment damage and product quality risks virtually but efficiently. Besides, for deep learning vision applications, the system is also better fitted to the factory environment for data augmentation and optimization of neural network models.