基于图像处理与信息融合的牛乳体细胞快速检测关键技术研究

Somatic cell count in milk is an important factor of mastitis detection and milk quality evaluation. Direct screening test method based on image processing is the standard method of somatic cell count in milk, and the key is somatic cell image segmentation. The image segmentation results will directly affect the accuracy and processing speed in real time. So considering the intelligent screening test method of somatic cell count as the study objective, using microscope images of milk samples collected in Inner Mongolia area, we further study somatic cell image segmentation and overlapping cell image separations. Because the image segmentation methods of 3-D color spaces consume large amount of computation and have slow speed, they do not suit to real-time applications. For those reasons, an improved Watershed approach in 2 dimensions of color spaces is studied, and the 3-D histogram will be reduced. The improved watershed segmentation methods are implemented in several lower color spaces, and then all segmentation results will be fusioned. The new segmentation models and the models of quantitative evaluation for the segmentation will be established. The overlapping cells are separated using generalized distance in order to resolve the problem of un-segmentation and over-segmentation. Finally the somatic cell count test method fitting for large sight images is formed and the theoretic bases of intelligent screening somatic cell count test system is established.

牛乳体细胞数量是判断奶牛乳腺炎、牛奶质量的重要指标。基于显微图像处理技术的直接镜检方法是牛乳体细胞数检测的标准方法，其关键是体细胞图像分割，分割方法直接影响体细胞数检测准确性及实时性。为此，以研究乳业生产中牛乳体细胞数智能镜检方法为目标，针对采集于内蒙古地区牛乳乳样的显微图像，进一步研究牛乳体细胞图像分割、分离等相关技术。鉴于3维彩色空间的图像处理运算量大，速度慢，拟研究基于2维直方图的彩色Watershed改进算法，对牛乳体细胞3维直方图降维，将几个低维空间分割结果实施融合，不降低分割准确度，提高分割速度10倍以上，形成新的牛乳体细胞分割模型，并定量评价；针对分割后牛乳体细胞在空间分布上相互重叠的特征，结合彩色图像空间统计特性的广义距离，研究重叠牛乳体细胞分离模型，并解决重叠牛乳体细胞未分割及过分割问题，最终形成适合大视野图像的牛乳体细胞快速检测方法，为建立智能化镜检检测系统奠定基础。

牛乳体细胞数量是判断奶牛乳腺炎、牛奶质量的重要指标。结合动物病理学，采用图像处理技术与信息融合技术，研究彩色牛乳体细胞图像的分割与评价方法及重叠体细胞分离方法，实现快速的牛乳体细胞数检测，为标准化的实时牛乳体细胞显微图像检测分析奠定理论与应用基础。研究选取RGB空间作为彩色牛乳体细胞图像处理的基础，利用基于不确定理论的云模型和基于降维与融合的方法对图像进行分割。对于粘连牛乳体细胞图像采用Hough圆检测的算法实现分离。在此基础上，提取基于Gabor-based(2D)2PCA频域整体特征，利用决策层融合策略完成整体与局部特征的融合，使用最近邻分类器识别，其识别率达98.5%，实验结果验证了算法的精度和稳定性。