Research On Key Techniques Of Multi-class Image Classification

With the rapid development of Internet and Mobile Terminal technologies,more and more users are willing to upload their images to the Internet.However,most uploaded images are disorganized.Although they can be well sorted manually,the labor cost is inestimable.Therefore,how to effectively use computer vision techniques in order to classify them becomes a critical issue for both academia and industry.Image classification is a typical pattern recognition problem.It can be categorized into two-class image classification and multi-class image classification according to the amount of classes involved.Obviously,multi-class image classification has more widely real-world applications and greater practical value.Therefore,image classification is generally referred to as multi-class image classification.Although many existing machine learning approaches can be used to solve multi-class classification problem,they rarely take into consideration the characteristics of multi-class image classification itself.For example,the total number of the classes can be very large,the difference among different classes can be very small,some classes can be mutually non-exclusive,and some target classes may not have any sample.In this paper,a series of research work has been carried out according to these characteristics,and the main contributions of this paper are summarized as follows:1)A selective subspace learning method is proposed to solve multi-class image classification problem in the case that the quantity of the class is relatively large.Although conventional supervised subspace learning methods are able to resolve the conflict between the high-dimensional feature problem and the small sample size problem to some degree,they fail to consistently improve image classification performance when the number of classes grows large.To tackle this issue,we propose a selective subspace learning method,which conducts local subspace learning for every testing sample by mining the sample-class relationship.This help to resolve the challenge that conventional subspace learning approaches encounter and therefore enhances the image classification task.2)An active annotation set based deep learning method is proposed for fine-grained image classification.In the fine-grained scenario,the inter-class difference is usually very subtle,i.e.,a category is easily confused with some other categories and such confusion usually makes it difficult to learn an effective classification model.In this paper,we analyze the inter-class confusion problem and propose an active set based deep learning method aiming at tackling those confusing classes.The proposed deep classification model has been demonstrated more effective than general deep networks.3)A novel deep learning framework based on random crop pooling is proposed for multi-label image classification.In multi-label images,the object layouts and scenes are usually very complex,which brings great difficulty for the training of the classifiers.In this paper,we propose a random crop pooling based deep learning method,which is able to discover the informative regions of the multi-label images automatically.Therefore,it is able to eliminate the background noise and improve multi-label classification performance.4)A unified framework of attribute regression and class prototype rectification is proposed for zero-shot image classification.In the zero-shot setting,no testing images are given in the training phase.Therefore,one needs to collect intermedia-level information like attributes to connect the training classes and the testing ones.The commonly used attribute regression methods are able to make such connections to some degree.However,they easily introduce projection domain shift problem and hubness problem.To tackle these problems,we propose a new approach by performing attribute regression and class prototype rectification jointly.