基于显著性感知原理的匀微纹理表面缺陷检测机器人视觉计算模型与检测算法研究

For a lot of manufacturing process, online surface defect inspection is an important part of product quality control. The key of even micro textured surface defect inspection robot application is to solve the challenging problem of vision: how to make defect inspection can be done with high accuracy and high speed , and make the algorithm be adaptive enough. To solve the difficult problem, this research is inspired by human eye inspection process. To lay the foundation for the establishment of high accuracy visual saliency map model, this project researches on the mechanism of calculation perceive for even micro texture surface defect. This project builds the models to imitate the function of retina and visual cortex V1 region, the model to compute visual saliency map. This project also develops the method for determining defects based on the characteristics of visual saliency map energy distribution, the method for segmenting defect goals; hence the inspection algorithm is adaptive. Heterogeneous computing can make bionic vision computation process be in accordance with both features of parallel processing and hierarchical serial processing in human visual system. Both the across-scale detection strategy of small defect and the algorithms based on heterogeneous computing are researched to enhance the the computing speed of the inspection process. New visual computing models, inspection algorithms will be proposed and are benefit to lay the theoretical foundation for the relevant applications.

产品表面缺陷的在线检测对于很多生产过程都是产品质量控制的重要环节，匀微纹理表面缺陷检测机器人的应用关键在于解决视觉方面的挑战性难题：如何使缺陷检测同时具备高准确率和高实时性，并且使算法有较强的适应性。针对该难题，本项目受人眼检测过程的启发开展研究。研究计算感知匀微纹理表面缺陷视觉显著性的机理，为建立高准确性的视觉显著图计算模型奠定基础；建立模拟视网膜及视皮层V1区功能的计算模型和视觉显著图计算模型；研究基于视觉显著图能量分布特征的缺陷判定方法及缺陷目标的分割方法，使检测算法具有适应性强的特点；异构计算可以使仿生视觉计算过程更符合人类视觉系统既并行又分级串行的特征，本项目研究微小缺陷跨尺度检测策略和基于异构计算的算法以提高检测过程的计算速度。本项目将在匀微纹理表面缺陷检测机器人视觉计算模型和检测算法方面取得突破，并为相关应用奠定理论基础。

产品表面缺陷的在线检测对于很多生产过程都是产品质量控制的重要环节，采用机器代替人进行表面缺陷检测成为趋势。本项目基于视觉显著性感知原理，开展了匀微纹理表面缺陷视觉计算模型和检测方法的研究，并研制了样机系统。.在项目实施中，建立了匀微纹理表面缺陷图像数据库，数据库包含了多种对象的典型缺陷样本；研究了计算感知匀微纹理表面缺陷视觉显著性的机理，明确了多模型组合和轮廓特征对于预测视觉显著目标的作用；研究了视觉显著性感知计算模型、跨尺度检测策略，提出了基于轮廓、亮度、颜色特征与跨尺度融合计算表面图像视觉显著图的算法；提出了分割表面图像的方法及缺陷判定方法；研究了基于深度学习和贝叶斯小样本学习的表面缺陷检测方法，提出的方法分别 在准确性及缺陷样本数量要求方面有更佳表现；研发了表面缺陷检测软件、布匹瑕疵检测系统样机、导光板缺陷检测系统样机、竹条缺陷检测机器人系统样机。.本项目提出的视觉算法与模型的准确性、计算效率、适应性均较好，为表面缺陷检测系统的研制奠定了理论基础。