Research On The Application Of Multi-instance Learning In Computer Vision

Multi-Instance Learning(MIL)is a special kind of classification problem where samples(called instances)are grouped into bags and labels are given only on the bag level instead of the instance level.Researchers pay increasing attention on this issue,and many inviting problems such as image classification,video annotation and object detection can be formulated in MIL frameworks.In this thesis,we study the MIL problem with two computer vision tasks,image classification and retrieval and abnormal event detection in videos.First,we propose an instance selection based MIL method.To address the image classification and retrieval,we improve the algorithm by utilizing a clustering strategy and multiview features.For the second task,abnormal event detection,we propose a novel method which is the combination of a graph-based MIL model and an improved dictionary learning.Finally,we implement the algorithms and make analyses according to the experimental results.The main technical contributions and novelties are:1)In this thesis,we combine the instance selection based MIL model with clustering and propose a novel MIL model called Clustering-based MIL(CMIL).Instance selection which is also called instance reasoning,is a significant step in the model,because the selected instances will be regarded as the positive samples when we train the final classifier.Considering the unlabeled data,clustering methods fit well for obtaining the hidden structure information in the feature space,and thus we utilize a clustering-based strategy to exploit such information on discovering positive instances.Instances in positive bags will be clustered into several groups,which makes these instances carry implicit structure information and keeps inter-bag instances independent.After that,we build up similarity graphs by three different measurements,direct similarity,discrinative similarity and consistency of similarity.Based on graphs,we select the positive instances.In addition,to fully use the original input data,we further develop our method in multiview feature space.We embed the multi-view features in both instance selection and classifier training.We implement our method in a parallel manner,which greatly reduces the time cost and enables the program to handle the large-scale data.2)In this thesis,we propose an Anomaly-introduced Learning(AL)to detect anomalies in videos.Preivous studies give many methods to address the issue,but there remains some challenges like too much time cost and unsatisfactor performance of locating the anomaly.To solve these problems,a graph-based MIL model is formed with both normal and abnormal video data.The MIL model generates a set of potentially abnormal instances and a coarse classifier.These instances are adopted for an improved dictionary learning method which we call anchor dictionary learning(ADL).The sparse reconstruction cost(SRC)is selected to measure the abnormality.Compared with other methods,we make use of abnormal information and prune testing instances with a coarse filter,and reduce time cost of computing SRC.Experiments demonstrate the effect of our proposed AL method by competitive performance.