Research On MIL-Ensemble Learning Theory And Its Application

MIL is the fourth class of machine learning framework which following the supervised learning, unsupervised learning and reinforcement learning. Different from previous learning frameworks, samples and instances are different concept in the MIL training data. In recent years, MIL gains high attention from scholars and has become a research hot spot as a new learning pattern. MIL has been widely applied in machine vision, image processing, etc. This paper studies MIL systematically and combines it with ensemble learning. The MIL-ensemble learning theory and methods are applied into object tracking and CAPTCHA recognition after deeply exploring it. The detail works in this paper can be listed as follows:(1) Firstly, this paper deeply researches the theories and principles of MIL-ensemble learning algorithm. Secondly, this article introduce three representative MILBoost algorithms. And then expounds respectively advantage and disadvantage of the application of three algorithms on object detection, tracking and recognition. All of the work is preparation for later target tracking and CAPTCHA recognition.(2) For the positive instances in positive bags may not distinctive when adaptively extracting samples in object tracking using MIL algorithm. It is easily be disturbed by useless or harmful instances, and is ineffectively in extracting the discriminative object features, this affects the robustness of the algorithm. A new weight instances selection algorithm based on kernel density estimation is proposed. Firstly, according to the weights are set according to the samples’distances from the object, that is,the sample which is closer to object has greater weight, otherwise smaller. Secondly, a kernel density estimation function is built to optimized positive bags and removed the useless and harmful instances in training set. Finally, optimized samples are used to train and learn, it is applied to realize real time tracking.(3)For dynamic CPATCHA designed by a website which has various types and big difference in shapes and sizes, and introduces time dimension to increase the difficulty for recognition. A new algorithm named MILBoost recognizing dynamic CPATCHA is proposed. However, misrecognition happens for partial matching when recognizing CPATCHA. To solve this problem and reduce misrecognition rate, overlap area comparison is used to recognize, and successfully recognize different types of this website’s dynamic CPATCHA. The proposed algorithm can achieve a high recognition rate on varies CPATCHA.