在TFT的制程中会造成许多不可避免的工艺缺陷, 需要逐个判断其对TFT电路所造成的影响, 这种缺陷检查方式需要大量人力且速度慢、精度低, 因此, 将人工识别替换成计算机自动化识别就显得尤为重要。自动识别中的一个关键部分就是将TFT电路进行定位。针对因缺陷颜色、形状、位置、大小不固定等原因导致现有图像处理算法无法准确定位出TFT电路位置的问题, 本文提出基于自相关性和模版匹配的TFT电路图像重构算法。首先, 对全自动光学检测(AOI)相机采集的缺陷图片进行区域提取, 估算出偏角并校正图像; 根据边缘图像的灰度垂直响应的自相关估计函数来估算电路的重复周期, 再由周期灰度均值来确定一组的电路纵贯线位置, 利用TFT电路周期特征重构出其余纵贯线; 通过角模版匹配算法重构TFT电路硅岛部分。实验结果表明, 本文提出的算法能够重构出被缺陷遮挡、图片模糊、电路缺失等情况下的电路, 定位准确率达96%以上。基本满足TFT缺陷自动化识别中的电路定位要求。

The TFT manufacturing process can inevitably result in a number of process defects, it is necessary to individually determine the impact caused by the defects. This defect inspection method requires a lot of manpower and slow, low accuracy. Therefore, it is especially important to replace manual recognition with computerized automatic identification. A key part in automatic identification is to locate the TFT circuit. Aiming at the problem that the existing image processing algorithms can not accurately locate the position of the TFT circuit due to the defect color, shape, position and size unfixed. This paper proposes an image reconstruction algorithm based on autocorrelation and template matching. Firstly, the region extraction is performed on the defect image acquired by the automated optical inspection(AOI) camera, and estimates the angle and corrects the image. The circuits repetition period is estimated according to the autocorrelation estimation function of the gray-scale vertical response of the edge image, and then the periodic gray-scale mean is used to determine the position of a group of longitudinal lines, and the remaining vertical lines are reconstructed by using the periodic characteristics of the TFT circuit. Finally, the silicon island part of the TFT circuit is reconstructed by the angular template matching algorithm. The experimental results show that the proposed algorithm can reconstruct the circuit under the condition of defect occlusion, blurred picture, and missing circuit. The accuracy of positioning is over 96%. Basically, it can meet the circuit segmentation requirements in TFT defect auto-identification.