Research On The Image Analysis Method Based On Multiple Instance Learning

Multiple instance learning is considered as the fourth category of machine learningmethods, which following supervised learning, unsupervised learning and reinforcementlearning. It brings new vitality to machine learning and has been widely applied in the fieldof image processing, stock market analysis. In this paper, we research on the basic theoryand algorithm of multi-instance learning, and applied it to the object recognition andtracking in the field of image analysis, The main works of the dissertation can be organizedas follows:1. We have summarized two aspects from the multi-instance learning theory, one ismulti-instance learning algorithm research, multi-instance learning was first proposed fromthe musk molecular classification problem. After that, multiple instance learning waswidely concerned by the machine learning community; scholars develop the new algorithmto solve the multi-instance learning problems. This article focuses here on the DDalgorithm and its significance. The other is multi-instance learning applications research.Multi-instance learning has been widely used in stock recommendations, text classification,image retrieval and many other fields.2. For traditional machine learning difficulties to get samples in the characterrecognition, we propose a multi-instance learning algorithm that base on mean-shift+support vector machines (Support Vector Machine, SVM). The algorithm take the pictureas a bag, extract local feature of the picture as instance of a bag; then using improvedMean Shift algorithm adaptively digging a characterization of the concept of the bagmarked instance, and remove interference instance of the bag. Finally using the SVMalgorithm to learning reduced bags. Last we built a multi-instance multi-valued SVMclassifier system to recognize characters.3. For traditional online learning object tracking method prone to drift in the complexenvironment, we proposed an online multiple instance object tracking method based onimproved random forest. We have constructed a multiple instance learning classifier whichcan be able to distinguish between the background and object, it package several imagesub-blocks which around target a certain area as positive bag of instances, so it has betterfault tolerance. We also improve random forest learning method, so that it can deal withmulti-instance learning problems, the system get positive bags and negative bags in an on line way, The results show it can respond to the changes of object’s shape, light andbackground.