医药大输液微小异物视觉检测与识别方法研究

Medical infusion solution is one of the five important preparations of the pharmaceutical industry in China. During the process of pharmaceutical manufacturing, some tiny insoluble foreign substances such as glass chips, hairs, fibers appear in the liquid medicine due to incomplete production and packaging technology, which can cause great hurt or even death. However the tiny insoluble foreign substances is versatile, no predefined shape, color and texture features, poor intensity contrast, Low SNR as well as combining complex noise and disturbance, it is a big challenge to inspect and recognize the tiny insoluble foreign substances automatically. Consequently, the main research contents of this project are as follows:.1) Reveal the difference between foreign substances, noises or disturbances, image sequences aligning,inspection region selecting, adaptive total variation model image denoising based on differential curvature are proposed to reduce computational complexity and improve the running efficiency..2) Research on insoluble foreign substances segmentation with improved fuzzy cellular neural network(IFCNN), feature selection and extraction algorithms based on injection image sequences..3) Research on BRIEF,BRISK,SURF and SIFT local feature description operator to extract and descript feature stably and effectively. Derive a novel approach based on M-estimators of Online Sequential Extreme Learning Machine called RRB-OS-ELM for identify multiple tiny insoluble foreign substances..The project applies research results will solve the infusion solution image preprocessing and inspection challenges,improve the imaging processing speed and precision,provide theoretical basis for the small moving target detection in the liquid.

医药大输液是医药行业五大重要制剂之一，由于生产工艺及灌装技术等原因，大输液药品中可能含有玻璃屑、毛发、漂浮物等可见异物，异物的存在严重危害用药安全。针对异物微小、种类多样、无固定形状和纹理特征、工业现场噪声干扰大等条件下的异物目标检测识别难题，重点研究：（1）揭示复杂工况背景下运动微小异物目标与噪声干扰的差异性，提出序列图像配准、检测区域标定、基于差分曲率的自适应全变分图像噪声抑制预处理方法，提高后续处理准确率；（2）研究药液图像中微小异物目标改进模糊细胞神经网络分割方法，快速准确提取运动异物；（3）研究BRIEF、BRISK、SURF和SIFT等快速局部特征描述算子，稳定有效提取和描述特征，设计基于稳健估计的在线序贯极限学习机识别分类算法。项目研究成果将解决复杂制药生产环境下液体运动目标的图像检测识别等技术难题，提高医药异物在线检测的处理速度和精度，为液体中运动小目标的检测识别提供依据。

医药大输液是医药行业五大重要制剂之一，由于生产工艺及灌装技术等原因，大输液药品中可能含有玻璃屑、毛发、漂浮物等可见异物，异物的存在严重危害用药安全。针对异物目标微小、种类多样、无固定形状和纹理特征、工业现场噪声干扰大等条件下的异物目标检测识别难题，重点研究并完成了：1）揭示复杂工况背景下运动微小异物目标与噪声干扰的差异性，提出序列图像配准、检测区域标定、基于差分曲率的自适应全变分图像噪声抑制预处理方法，提高后续处理准确率；2）研究药液图像中微小异物目标改进模糊细胞神经网络分割方法，快速准确提取运动异物；3）研究BRIEF、BRISK、SURF和SIFT等快速局部特征描述算子，稳定有效提取和描述特征，设计了改进ELM算法和基于改进联合稀疏表示的运动异物目标识别分类算法，项目成果录用IEEE Transactions on instrumentation and measurement 2篇，发表SCI 收录论文 1 篇，EI 收录论文 5 篇，会议论文 4 篇，授权中国发明专利 3 项，申请中国发明专利4项。获省部级奖励3项。项目负责人入选湖南省首届“湖湘青年英才”。项目研究成果解决了复杂制药生产环境下液体运动目标的图像检测识别等技术难题，提高医药异物在线检测的处理速度和精度，为液体中运动小目标的检测识别提供依据，项目提出的多种检测算法识别率高，对工程设备的研制具有重要意义。