Research On Object Counting Based On Support Vector Machine And Regularized Risk Minimization

With the rapid development of technology and theory, intelligent image and video analysis has become a very active area of research, it is mainly to study the images and video sequences of interest in target detection, tracking, recognition and behavior analysis. As important research area of intelligent image and video analysis, more and more attention has been paid to density estimation and object counting.This thesis researches two kinds of object counting methods, and uses urine sediment microscopy images to verify the feasibility of these methods. Firstly, we studied the urinary sediment image segmentation as a prerequisite. The following study is divided into two aspects:For low-density multi-class object groups, we studied a method for object classification and counting based on SVM. We studied the machine learning method using support vector machine, and introduced it to the study of classification and counting of the visible components of the urinary sediment. Choose image features that can well characterize different cell types in the urine sediment microscopy image through experiments, extract them to generate the feature vector, then use two cascading SVM to successively classify red and white blood cells. Once classified, the count of each type of cells is readily known.For high-density object groups, we proposed a method for object counting based on regularization risk minimization. We introduced the definition of the object density function, whose integral over any image region gives the count within that region. Then we proposed an approach to construct ground truth density function, and derived a parametric model of the density function. Lastly, estimate the model parameter based on the principle of regularization risk minimization, and cast this minimization problem into a linear programming problem. Unlike common approaches, for every input image we estimate a density function, whose integral over any image region gives the count within that region. For high density object groups, our approach also gives counts, not only density levels.