Zero-shot Image Classification Based On Deep Learning And Knowledge Mining

As labeled training samples are inadequate to cover all the object classes in zero-shot image classification, it is impossible for traditional classifiers to recognize new object classes emerging at the testing stage. For the attribute-based methods aiming at resolving the zero-shot image classification, attributes serve as an intermediate of knowledge transfer connecting the visible and unknown new modes.Existing zero-shot learning methods mainly have several problems as follows. Various prior knowledge associated with attributes fails to be fully utilized in designing attribute learning model; In the attribute-category mapping, the classification capability description of attribute classifiers is insufficient; The dependency between low-level features and high-level classes is not fully depicted, etc. Based on some of these problems, the thesis is arranged as follows:Firstly, image features used in the attribute-based zero-shot learning methods are subjected to manual extraction, in which the classification accuracy depends highly on quality of the feature to be extracted. Based on this problem, a novel zero-shot image classification method based on deep feature exaction is proposed(DLIAP). First,image patch extraction and ZCA whitening are applied to reduce the computational complexity and correlations between pixels; Then, the image features are learnt through a stacked sparse auto-encoder to obtain a feature mapping matrix, and further the feature mapping matrix is proposed as a convolution kernel to convolve and pool with images; Finally, the exacted image features are used on the zero-shot learning.Secondly, at present the zero-shot learning about training attribute classifiers still rely on manual feature extraction and the shallow learning model, and the description of each attribute classifier is insufficient. To solve this problem, a novel zero-shot learning model is proposed by deep weighted attribute prediction(DWAP). The algorithm uses supervised learning method for training a deep convolution neural network and adds the attribute label into the network to achieve a deep feature presentation and attributes prediction. Further, by means of mining prior knowledge of attributes, attributes classifiers of different expression power are performed with weighting design. The algorithm constructs a direct weighting attribute prediction model and yields more accurate attribute prediction and zero-shot image classification.Thirdly, the existing attribute-based zero-shot learning models at different levelslacks some necessary prior knowledge. To solve this problem, a novel zero-shot learning model is proposed by exploiting class-related and attribute-related prior knowledge(IAP_CAPK). First, Whitened Cosine Similarity(WCS) is applied to discover class-class correlations that are used to screen out training samples having high correlations with testing samples; Then, Sparse Representation Coefficient(SRC)is used to mine attribute-related prior knowledge and to discover attribute-class and attribute-attribute correlations. Such a couple of correlations are respectively used to screen out attributes having high correlations with testing classes and also to remove redundant attributes; Finally, training sets of strong relativity to testing classes and attributes are used on the zero-shot learning.Experimental results on several databases show that, when compared with several widely-used zero-shot image classification methods, the proposed methods can yield more accurate attribute prediction and better zero-shot image classification.