基于多算法融合Mask R-CNN堆叠千克组砝码识别与关键部位分割方法  

Image instance segmentation technology for critical parts of stacked kilogram weights

作　　者：赵迪 刘桂雄[1] 马健 郭琳琳 ZHAO Di;LIU Guixiong;MA Jian;GUO Linlin(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China;Guangzhou Institute of Metrology and Testing Technology,Guangzhou 510030,China)

机构地区：[1]华南理工大学机械与汽车工程学院,广州510640 [2]广州计量检测技术研究院,广州510030

出　　处：《激光杂志》2021年第5期27-31,共5页Laser Journal

基　　金：国家市场监督管理总局科技计划项目(No.2019MK086)。

摘　　要：砝码无人化检定有助于提高砝码检定效率、增强检测数据可溯源性。从深度学习实例分割方法入手,提出用于砝码无人化检定的基于Mask R-CNN堆叠千克组砝码识别与分割技术方案,通过数据增强、图像对比度增强、多算法融合等技术方案实现日光灯照明下堆叠千克组砝码图像关键部位识别与分割,形成泛用性强的人工智能方法,通过模拟实验获得较为合适的训练集图像增强算法与测试集图像增强算法,实现图像中实例100%识别,与原始Mask R-CNN相比AP_(50)提升21.51%,Mask IoU提升17.95%。Self-service verification of weights helps improve the efficiency of weight verification and enhance the traceability of test data. This paper starts with the deep learning instance segmentation method and proposes a technology solution to recognise and segmentation stacked kilogram weights based on Mask R-CNN. Using technology such as data enhancement,image contrast enhancement and algorithm fusion,which realised the identification and segmentation of critical parts of the stacked kilogram weights under fluorescent lighting conditions. They formed a highly versatile artificial intelligence method. The simulation experiments found the most suitable training set image enhancement algorithm and test set image enhancement algorithm. The instances in the image are 100% identified. Compared with the traditional Mask R-CNN,the AP_(50) has increased by 21. 51%,and the segmentation integrity Mask IoU has increased by 17. 95%.

关 键 词：千克组砝码 低对比度识别 图像对比度增强 实例分割 深度学习 

分 类 号：TN249[电子电信—物理电子学]