摘要
为了实现CTP版材表面不同类型瑕疵的自适应在线检测,引入人类视觉注意机制,融合区域生长算法对表面瑕疵图像进行分割,以面积作为特征值对瑕疵进行识别。首先根据人类视觉系统具有选择性注意机制的特性,对CTP版材表面图像的多特征图像采用图像金字塔分解得到多尺度图像,并利用基于图的算法(GBVS)将多尺度下多特征的图像融合为全局显著图。然后将显著图中最显著点即注意焦点作为区域生长的种子点,相似性作为区域生长准则,对原图像的梯度图像采用区域生长算法获得瑕疵区域二值化图像。最后利用注意抑制返回机制和邻近优先的准则,查找其他未检瑕疵,直至满足CTP版材的行业标准。实验结果表明:瑕疵检测系统分辨率0.1mm,检测平均准确率达96.3%以上,算法运行速度快,能满足CTP版材生产的在线检测实时性要求。
In order to realize automatic detection for flaw on CTP plate,an automatic detection system is established and its applied algorithms such as region growing and graph-based visual saliency（GBVS） is investigated.First,according to biological characteristics of the human visual attention mechanism,image elementary features were extracted by sampling the center-surround differences,which were combined into a saliency map.Then competition among salient points in this map gave rise to a single focus of attention（FOA） which was selected as the seed point of region growing segmentation.Taking similarity for region growing criteria,binary image of flaw region was formed.Finally,after computing characteristic parameters according to flaw region and using area features,flaw on CTP plate surface were identified.Experimental results indicate that intelligent detection of flaw on CTP plate achieves system resolution of0.1mm with average accuracy rate of 96.3%.It can satisfy the system requirements of non-contact,online,real time,higher precision,rapid speed and stabilization.
出处
《长春理工大学学报（自然科学版）》
2014年第6期94-98,共5页
Journal of Changchun University of Science and Technology(Natural Science Edition)
关键词
机器视觉
表面瑕疵
视觉注意
显著图
区域生长
machine vision
surface flaw
visual attention
saliency map
region growing
