Cross-modal hashing has attracted much attention due to its low storage cost and high retrieval efficiency. Most of the existing cross-modal hashing methods require the inter-instance association information provided by additional manual labels. However, the deep features learned by pre-trained deep unsupervised cross-modal hashing methods can also provide similar information. In addition, the discrete constraints are relaxed in the learning process of Hash codes, resulting in a large quantization loss. To solve the above two issues, a Deep Unsupervised Discrete Cross-modal Hashing (DUDCH) method based on knowledge distillation was proposed. Firstly, combined with the idea of knowledge transfer in knowledge distillation, the latent association information of the pre-trained unsupervised teacher model was used to reconstruct the symmetric similarity matrix, so as to replace the manual labels to help the supervised student method training. Secondly, the Discrete Cyclic Coordinate descent (DCC) was adopted to update the discrete Hash codes iteratively, thereby reducing the quantization loss between the real-value Hash codes learned by neural network and the discrete Hash codes. Finally, with the end-to-end neural network adopted as teacher model and the asymmetric neural network constructed as student model, the time complexity of the combination model was reduced. Experimental results on two commonly used benchmark datasets MIRFLICKR-25K and NUS-WIDE show that compared with Deep Joint-Semantics Reconstructing Hashing (DJSRH), the proposed method has the mean Average Precision (mAP) in image-to-text/text-to-image tasks increased by 2.83 percentage points/0.70 percentage points and 6.53 percentage points/3.95 percentage points averagely and respectively, proving its effectiveness in large-scale cross-modal retrieval.

Unsupervised feature selection attracts much attention in the field of machine learning, and is very important for dimensionality reduction and classification of high-dimensional data. The similarity between data points can be measured by several different criteria, which results in the inconsistency of the similarity measure criteria between different data points. At the same time, in existing methods, the similarity matrices are most obtained by allocation of neighbors, so that the number of the connected components is usually not ideal. To address the two problems, a Self-Adaptive Multi-measure unsupervised feature selection with Structured Graph Optimization (SAM-SGO) method was proposed with regarding the similarity matrix as a variable instead of a preset thing. By fusing different measure functions into a unified measure adaptively, various measure methods could be synthesized, the similarity matrix of data was obtained adaptively, and the relationships between data points were captured more accurately. In order to obtain an ideal graph structure, a constraint was imposed on the rank of similarity matrix to optimize the local structure of the graph and simplify the calculation. In addition, the graph based dimensionality reduction problem was incorporated into the proposed adaptive multi-measure problem, and the sparsity-inducing l _2,0 regularization constraint was introduced to obtain the sparse projection used for feature selection. Experiments on several standard datasets demonstrate the effectiveness of SAM-SGO. Compared with Local Learning-based Clustering Feature Selection (LLCFS), Dependence Guided Unsupervised Feature Selection (DGUFS) and Structured Optimal Graph Feature Selection (SOGFS) methods proposed in recent years, the clustering accuracy of this method is improved by about 3.6 percentage points averagely.

To solve the problems that group effect is neglected when selecting dictionary bases in sparse coding models, and distance between a features and a dictionary base can not be effectively measured by Euclidean distance, Non-negative Local Sparse Coding algorithm based on Elastic net and Histogram intersection (EH-NLSC) was proposed. Firstly, with elastic-net model introduced in the optimization function to remove the restriction on selected number of dictionary bases, multiple groups of correlation features were selected and redundant features were eliminated, improving the discriminability and effectiveness of the coding. Then, histogram intersection was introduced in the locality constraint of the coding, and the distance between the feature and the dictionary base was redefined to ensure that similar features share their local bases. Finally, multi-class linear Support Vector Machine (SVM) was adopted to realize image classification. The experimental results on four public datasets show that compared with LLC (Locality-constrained Linear Coding for image classification) and NENSC (Non-negative Elastic Net Sparse Coding), the classification accuracy of EH-NLSC is increased by 10 percentage points and 9 percentage points respectively on average, proving its effectiveness in image representation and classification.

Focused on that limitation of single-layer structure on image feature learning ability, a deep architecture based on sparse representation of image blocks was proposed, namely Multi-layer incorporating Locality and non-negativity Laplacian Sparse Coding method (MLLSC). Each image was divided uniformly into blocks and SIFT (Scale-Invariant Feature Transform) feature extraction on each image block was performed. In the sparse coding stage, locality and non-negativity were added in the Laplacian sparse coding optimization function, dictionary learning and sparse coding were conducted at the first and second levels, respectively. To remove redundant features, Principal Component Analysis (PCA) dimensionality reduction was performed before the second layer of sparse coding. And finally, multi-class linear SVM (Support Vector Machine) was adopted for image classification. The experimental results on four standard datasets show that MLLSC has efficient feature expression ability, and it can capture deeper feature information of images. Compared with the single-layer algorithms, the accuracy of the proposed algorithm is improved by 3% to 13%; compared with the multi-layer sparse coding algorithms, the accuracy of the proposed algorithm is improved by 1% to 2.3%. The effects of different parameters were illustrated, which fully demonstrate the effectiveness of the proposed algorithm in image classification.

Most of the semi-supervised multi-view feature reduction methods do not take into account of the differences in feature projections among different views, and it is not able to avoid the effects of noise and other unrelated features because of the lack of sparse constraints on the low-dimensional matrix after dimension reduction. In order to solve the two problems, a new Semi-Supervised Adaptive Multi-View Embedding method for feature dimension reduction (SS-AMVE) was proposed. Firstly, the projection was extended from the same embedded matrix in a single view to different matrices in multi-view, and the global structure maintenance term was introduced. Then, the unlabeled data was embedded and projected by the unsupervised method, and the labeled data was linearly projected in combination with the classified discrimination information. Finally, the two types of multi-projection were mapped to a unified low-dimensional space, and the combined weight matrix was used to preserve the global structure, which largely eliminated the effects of noise and unrelated factors. The experimental results show that, the clustering accuracy of the proposed method is improved by about 9% on average. The proposed method can better preserve the correlation of features between multiple views, and capture more features with discriminative information.