Image Classification Based On Feature Representation And Coding

Image classification always plays an important role in machine learning,computer vision and pattern recognition.Image classification models usually consist of following steps:feature extraction,feature transformation and enhancement,classifier training.In order to obtain effective feature representation,early feature extraction mainly adopts manual de-sign,which usually relies on domain knowledge of feature designers.With the development of deep learning technology,more and more classification models use deep neural networks for feature extraction.The deep feature representation can be obtained through end-to-end training on deep networks,as well as directly obtained with pre-trained deep networks on large-scale image database.The former usually needs lots of training samples to fit network parameters,thus the calculation expense is high.In the latter,due to the lack of guidance of supervisory information,the discriminability of deep feature representation needs to be strengthened.As an effective feature transformation and enhancement technology,feature coding has made great strides in image classification.However,they still have some potential challenges in practical application,such as the number of categories is large but few training samples with labels,the image samples has large inter-class variance but small intra-class distance,the test classes have not training samples,the coding efficiency of image features among different levels is low,the stability of feature coding is poor.To deal with above chal-lenges,this dissertation based on feature coding technology with various of image features to conduct in-depth research on image classification.The main works and creations are as follows:(1)Based on multi-layer dictionary learning and feature coding,an effective image classifi-cation model is proposed that transforms the original handcrafted or deep network features into deep coding features.Compared with the methods based on single-layer dictionary learning and feature coding,the proposed model can reduce the impact of noise in the image data and enhance the robustness of dictionary atom,hence improve the classification perfor-mance on noisy images.The successive nonlinear transformations of original image features are realized via layer-by-layer dictionary learning,thus the separability of coding coefficient features from different classes is increased.Meanwhile,benefiting from the integration of supervised classifier training and multi-layer learning,the complex structure information in image data is exploited,and the discrimination of original image-level features is further improved.Besides,by utilizing the graph Laplacian matrices based on the dictionaries,very large values in the coding coefficient features can be avoided,which reduces the risk of over fitting and the classification error of test images.Compared with single-layer feature coding methods and deep network models,the proposed model can give consideration to the accura-cy and efficiency of image classification,and the experimental results on various benchmark image datasets demonstrate its notable classification performance(2)A method to generate coding features with class-wise sparse distribution is proposed,which can achieve better image classification.Focus on the generation mechanism of cod-ing coefficients for image classification,the proposed method first explores the class labels of training samples to construct a term of label information perception.By enhancing the coefficients of samples from same class and suppressing the coefficients of samples from different classes,this term can make the coding coefficients with class-wise sparse distribu-tion.Then the graph regularization term based on the structural information of samples is constructed,and integrated with label information perception term to make generated coding coefficients more smooth,also obtain a robust dictionary.Moreover,a support vector based classifier is introduced to make the coefficient vectors of different classes to be separated by a max-margin.The above constraint terms are interaction and strengthen in the learning process,finally transform the original image features into class-wise coding features which are more suitable for classification.Since all the constructed constraints are based on the l2 norm,it avoids time-consuming solution for solving l0 or l1 norm,therefore the proposed method can meet the application of real-time image classification.The experimental results show that the class-wise coding features generated by our method can enhance discrimina-tive ability of various original handcraft or deep network features(3)In view of inductive and transductive zero-shot image classification,this dissertation proposes two stacked semantic auto-encoder models respectively.In inductive classification situation,the test images from unseen classes are unavailable in the training stage,it is crucial to transfer the knowledge from the seen classes to unseen ones at this moment.To this end,this dissertation first proposes a model based on stacked semantic auto-encoder with manifold regularizations,which not only establishes the tight relations among the spaces of visual representation,semantic description and label information to narrow the semantic gap,but also fills the class-domain gap via integrating the manifold regularizers into basic model Compared with other zero-shot classification methods,the proposed model has stronger generalization ability,and can transfer the knowledge from the seen classes to unseen ones effectively.Aim at transductive classification situation where the test samples are available in the training stage,this dissertation further proposes a domain-aware stacked auto-encoder model,which consists of two parallel but interactive stacked auto-encoders where one is constructed for the training data from seen classes and the other for the testing data from unseen classes.Experimental results show that the proposed two models for zero-shot image classification based on encoder-decoder paradigm can achieve state-of-the-art performance under both the conventional and generalized zero-shot classification criteria.