Abstract: It is difficult to track the impurities, which land fast and are very similar in transfusion bottles. In this paper, an impurity detection method is proposed to solve this problem by embedding linear prediction and[FLBPriu28,1], which means five uniform texture models of Local Binary Pattern related to the edge and corner, into Mean shift algorithm. The background of image sequences is established using the simplified normalization cross correlation coefficients. The exact positions of the impurities are extracted applying background subtraction, gray image morphology, and maximum contrast segmentation. The improved Mean Shift is used to track motional impurities in a few successive frames to ensure detection accuracy. The experimental results show that the average detection rate is 96.3% for the impurity larger than two pixels in diameter. The average detection time is 0.8 seconds per transfusion bottle.

Key words: similarity, background subtraction, gray image morphological, LBP, Mean Shift, impurity automatically detection

摘要： 针对大输液中快速降落及相似杂质难以跟踪的问题，提出了一种将线性预测及[LBPriu28,1]纹理模型中表示边界和角的5种基本模式[FLBPriu28,1]（Five Local Binary Pattern）嵌入到Mean Shift算法中的异物检测方法，实现了对杂质目标的有效跟踪。利用简化的归一化互相关系数快速建立序列图像的背景，采用背景减除法、灰度图像形态学及最大对比度分割法提取目标杂质的精确位置，利用改进的Mean Shift算法连续跟踪数帧运动杂质确保检测准确率。实验结果表明，该方法对直径不小于3个像素的杂质检测率平均达到96.3%，检测速率平均达到0.8秒每瓶。

关键词: 相似度, 背景减除, 灰度形态学, LBP, Mean Shift, 异物自动检测