Research Of Polymorphous Bovine Somatic Cell Recognition Based On Feature Fusion

There are many methods to detect the number of bovine somatic cells.The California Mastitic Test(CMT),the pH value detection and the measurement of Electrical Conductivity(EC)are used successfully in industry.However,some methods only get a relative amount of the number of somatic cells instead of precise data.Some of them need long time to detect and tedious operation steps.Other equipments are expensive and need to be corrected frequently,which is only suitable for large quantities of samples to be measured together.When test results are abnormal,the most accurate and strict method based on microscopic image somatic cell count is used to retest.Because it is the standard method.According to the number of different types of somatic cells,it can further monitor the types of pathogenic bacteria,the degree of infection and the effect of mastitis on milk production.It is convenient to take corresponding treatment scheme in time to prevent disease spreading.Compared with blood cell images,although bovine somatic cells don't contain red blood cells,they contain a large number of milk fat,lactoprotein and debris.They will emerge noise interference for images and increase uncertainty factors during the image processing.In this dissertation,we take Holstein cows as the research objects and get somatic cell microscope pictures.Next,we do the pretreatment,image segmentation,feature parameter extraction and the establishment of somatic cell recognition model.The following conclusions have been made:(1)The optimal threshold of images is obtained by the proposed algorithm in this paper based on cloud model through the uncertainty theory.According to gray distribution characteristics of the nucleus,a new criterion is proposed to detect nuclei from the somatic cell image by the super entropy of cloud model.The criterion can excavate the target set more comprehensively and accurately.Maximizing the super entropy can relax the condition of the normal distribution and make the maximum tolerance of the gray value.This method not only finds the optimal threshold,but also effectively reduces the computational complexity.Experiments show that cloud model based on uncertainty theory is simple,effective and has strong anti-interference ability,which significantly improves the efficiency of cell image threshold.(2)Image segmentation is completed by the watershed algorithm based on distance transformation and seed points merging.First,the initial seed points are obtained through distance change.Due to the large difference in the shape,size and structure of somatic cells,obtained seed points are redundant.Then,we take full advantage of the spatial information of seed points,optimize seed points,reconstruct distance graph and achieve primary segmentation of images using fast invasive watershed algorithm.Finally,the advanced segmentation is completed by merging adjacent regions,which effectively suppresses defects of over segmentation.The integrated gray distance is used in regional merging,and the result is more similar to that of the expert's.The algorithm effectively suppresses the over segmentation phenomenon,and successfully segmented strongly adherent cells.The segmentation result is satisfactory,and the algorithm is not very strict in shape.(3)The overall characteristics of cells in frequency domain are obtained by Gabor-based(2D)~2PCA,the proposed cell recognition algorithm in this paper.Because Gabor filters have different scales and directions,they can decompose the image into more perfect frequency features,and effectively improve the robustness of the noise in the acquisition process.A high dimensional Gabor space is composed of Gabor features.The feature dimension of Gabor space is too high to handle directly by the computer.Therefore,in the two directions of row and column,dimensions are reduced simultaneously,and features cover the largest amount of information,which effectively reduces computational complexity and feature dimension.It not only improves the recognition efficiency,but also reduces the recognition time.The experiment shows that Gabor-based(2D)~2PCA algorithm is simple,but effective.Though the feature fusion,it has strong anti-interference ability and significantly improves the efficiency of cell image recognition.The decomposition characteristics of cells'spatial domain are extracted from the non negative matrix factorization.It can decompose the cell image into local parts,and get the important recognition information.(4)The fusion process based on global and local features is as follows:first,calculate the matching distance between the overall and local features.Second,the fusion process is to calculate the matching distance using the weighted sum method and the zero mean normalization method.Finally,we use nearest neighbor classifier to identify.The algorithm takes into account both the whole information of cell image in the frequency domain and the spatial domain information with a distinct local distinction.It has a certain complementarity.The recognition accuracy and stability of the system have been improved.Although the fusion strategy used in this algorithm is simple,it does not affect the efficiency of recognition.The experimental results verify the effectiveness and stability of the algorithm.(5)The optimal feature set obtains by Relief F and K-S test algorithm.First,the 4types of cell features are extracted separately,including color features,geometric features,invariant moments and texture features.According to the characteristics of Relief F algorithm,all weights are calculated.Then,the primary features are selected based on the predetermined threshold.K-S test is used to eliminate redundant features,and the optimal featurs are got.Finally the weighted fusion strategy is applied and we get the optimal fusion feature set.According to the performance of each feature in training set,we determine weight values.The better the performance is,the higher the weight value is.The optimal fusion feature set can independently determine the dimension of feature set,the accuracy of balanced classification,the elimination of redundant features,and the reduction of feature dimension.At the same time,the optimal fusion feature set reduces the feature dimension to the greatest degree,guarantees the workload of data preparation,improves operation efficiency and saves storage space.Therefore,the feature selection method proposed in this paper is a reliable implementation scheme,which is feasible and more suitable for the extraction of feature sets in somatic cell classification processing.