Research On Image Classification Technology Based On Multi-instance Learning

To solve the classification tasks,there are currently a variety of classification learning frameworks.Due to inaccurate or inadequate labeling of datasets in practical applications,the weakly-supervised learning framework is gradually being applied and developed.The multi-instance learning adopted in this paper belongs to the weakly-supervised learning framework.At present,the main multi-instance learning methods are generally divided into three categories: bag-space method,instance-space method and embedded-space method.However,There are some problems.On the one hand,at present,researchers pay little attention to the preprocessing of multi-instance learning — image bag generators.In order to overcome the shortcomings of k-meansSeg algorithm,this paper proposes an image bag generator called S-k-meansSeg which incorporates spatial information into k-meansSeg.On the other hand,this paper finds that the bag-space method is easy to ignore the local information in the bag,and the instance-space method is easy to ignore the overall structure information of the bag in current image classification methods based on multi-instance learning.Therefore,this paper proposes a multi-instance learning method to fuse bag-space features with instance-space features.The main contents of this paper are as follows:1.In this paper,we first investigate the current situation of multi-instance learning in China and other countries.And we expound the bag-space multi-instance learning method and the instance-space multi-instance learning method in detail and analyze their respective advantages and disadvantages.Then we investigate the related technologies of image classification based on multi-instance learning and contrast each related method.2.In this paper,the image bag generator — S-k-meansSeg which incorporates spatial information into k-meansSeg is proposed.For an image,the k-meansSeg method extracts the color and texture features of each image block.In this paper,we fuse the position feature with each image block features,and then all the image blocks are clustered.The clusters are instances of an image bag;all clusters make up an image bag.3.In this paper,a multi-instance learning method to fuse bag-space features with instance-space features is proposed for image classification.Firstly,we established a graph model that described structural relations among instances in a bag.The graph model was transformed as an affinity matrix which could be used as the bag-space features.Secondly,we selected the instances in the positive bags.The features of the instances would be regarded as the instance-space features,if the correlation between those instances and the category of the positive bag was relatively strong.And we selected the instances in the negative bags.The features of the instances would be regarded as the instance-space features of the negative bags,if the correlation between those instances and the category of the positive bag was weaker.Finally,we used the Gaussian RBF kernel to map the bag-space features and the instance-space features to the same feature space.Then we used the feature fusion method based on the weight to fuse the two kinds of features in the same feature space.The experimental results on multiple public image data sets show that the classification performance is improved by the proposed method.In addition,the multi-instance learning method,which fuses bag-space features with instance-space features is also applied to the task of drug molecule activity prediction and text classification,and has achieved good results.