摘要
在TFT-LCD导电粒子压痕检测过程中,由于检测对象面积小、数量多、分布不均等因素,传统的半导体检测方法漏检、误检较多,无法实现有效检测.为提高TFT-LCD导电粒子压痕的识别率和识别速度,提出一种基于图像梯度的TFT-LCD导电粒子快速检测方法.该方法针对导电粒子的阴阳面特征,提出了统计意义上的伪光照方向判定方法,解决目前光照模型不能适应导电粒子伪光照方向的判断,同时有助于后续梯度统一表述为阳面区域,然后提出了区域特征梯度的概念,有效避免导电粒子的常规梯度信息受噪声干扰严重的问题,并采用结合动态规划的K-means算法对区域特征梯度进行聚类计算,得出梯度阈值,完成对导电粒子压痕的检测与计数,算法识别率约为98.61%,漏检率为1.39%,误检率约为1.85%.此外,为了满足工业检测中对于亿级像素图像处理的实时性需求,采用基于CUDA平台的并行计算方法对算法进行了加速,核心算法部分加速约253倍,整体算法加速约8倍,代码运行时间<2 s,满足工业检测的实时性需求.
The traditional image threshold is not effective for detecting conductive particles in a thin-film transistor liquid crystal display(TFTLCD) circuit due to uneven illumination, micro size, and the large number of particles involved. To improve the accuracy and speed of detecting TFT-LCD conductive particles, a novel method based on the image gradient is proposed. Given the variable brightness of conductive particles in an image, a pseudo-illuminant-direction-detection estimation is presented. The estimation directly addresses the issue that the illuminant model method cannot be adapted to identify the pseudo-illuminant direction. Subsequently, the concept of a regional characteristic gradient is introduced. The regional characteristic gradient has the advantage of being less sensitive to noise.Furthermore, the K-means algorithm and modified K-means algorithms are applied to compute the threshold for detecting conductive particles. The accuracy of the proposed method reached 98.61%;the missing rate dropped to 1.39%, and the false rate was maintained at 1.85%. To meet real-time application needs, a compute-unified-device-architecture-based parallel computing method is adopted to speed up implementation. The run-time was significantly reduced using a core algorithm running 252.8 times faster, and the overall algorithm ran eight times faster than the conventional methods. The total implementing time of the proposed method was less than two seconds, thus satisfying real-time industry detection requirements.
作者
罗晨
范霆霄
张鑫
周怡君
贾磊
LUO Chen;FAN TingXiao;ZHANG Xin;ZHOU YiJun;JIA Lei(School of Mechanical Engineering,Southeast University,Nanjing 211189,China;Wuxi Shangshi Electronic Technology Co.LTD,Wuxi 214000,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2021年第2期231-240,共10页
Scientia Sinica(Technologica)
基金
国家自然科学基金(批准号:51975119,51575107)
“科技助力经济2020”重点专项资助项目。
关键词
导电粒子
图像处理
图像梯度
并行算法
聚类
conductive particles
image processing
image gradient
parallel algorithm
clustering
