Research On Image Feature Learning Algorithm And Its Applications

As an important source of visual information, image contains rich content and intuitive representation, thus it is regarded as the most commonly used information carrier in human activities. With the development of science and technology, driven by practical application requirements, image processing by using computers has become an important research area in visual analysis and machine learning. The key research problem of visual representation and learning is the construction of the accurate visual representation that better reflects the visual properties and semantic information.As the basic part of the image processing, the image representation capability has vital influence on the performance of image classification and image recognition. The research challenge of image representation is the complexity and diversity of the visual content brought by illumination, gesture and scene change. Based on the low-rank representation(LRR), sparse representation, nonnegative matrix factorization(NMF) and their combined usage, this thesis proposes several effective image representation methods for face images, breast histopathology images and some kinds of physical images respectively, and achieves the goal of improving the performance of image classification and recognition.The main contributions in this thesis are summarized as follows:In Section 2, a novel deep low-rank matrix decomposition method for image representations is proposed. Unlike traditional LRR methods that impose additional constraints on the representation and dictionary, the proposed approach iteratively performs low-rank decomposition to the coefficient matrix layer by layer, making the rank of the representation matrix to be lower and lower along the iterations. It can avoid the performance degradation problem in the image representation caused by the unreasonable parameter selection. First, under the conditions that the samples of the database are corrupted seriously, we build the first LRR module by setting a smaller value to the parameter that restricts the sparse noise, using the data collection itself as a dictionary and obtaining the representation matrix and noise. Then, we input the de-noised image as a dictionary to the second LRR module and conduct low-rank decomposition once again on the coefficient matrix in the first module. We repeat the low-rank decomposition operation until the rank of the coefficient matrix reaches an expected level. Compared to standard low rank representation learning approach, our deep low rank decomposition method achieves 16.43% higher recognition rate in visual recognition tasks on Extended Yale B dataset. The results prove that our method achieves better performance in image de-noising and classification on the serious corrupted database.Based on color deconvolution, in Section 3 we applies LRR, nonnegative sparse LRR and deep LRR(Section 2) to the representation learning of breast histopathology images, and three segmentation results are obtained. The results indicate that deep LRR approach has the best segmentation result. First of all, based on Beer-Lambert Law, images colored by Hematoxylin and Eosin(H&E) are transformed to optical density form. Then, by applying LRR, nonnegative sparse LRR and deep LRR approaches respectively, the H&E channels of H&E image are separated. Meanwhile, the resultant image of H component in RGB space is transformed into gray scale image and consequent binary image. Finally, the watershed algorithm based on mathematical morphology is employed to separate the connected nuclei, and nuclei segmentation process is performed by finding the nuclear contours. The above three methods can effectively separate the cells from the breast histopathology images, which provides better preparation to the subsequent quantitative analysis on the cells and also has great help for enhancing the accuracy of the final diagnosis to the cancer.By applying semi-supervised LRR and sparse matrix decomposition, in Section 4 we present a label constrained LRR(LCLRR) method, which incorporates the label information as an additional hard constraint for learning the pair-wise visual similarity in the visual graph, where each node represents an image. The proposed method applies the information in the labeled data and combines the intrinsic information in the remaining large number of unlabeled data. Accordingly, the combination of both labeled information and unlabeled information achieves better semantic consistency and better adaptation to the content divergence than using single information source. First, we impose constraints using label information to traditional LRR model. By approximating the objective function using a reasonable convex function, the convex function is optimized by augmented Lagrange algorithm. Then, by restricting the model with non-negativity and sparsity, the resulted image representation not only encodes both global visual structure and local visual characteristics of the image, but also embeds the semantic information in the labels. Finally, we construct an LCLRR graph by representing the current node as the linear combination of all other nodes and restricting the refactoring coefficient to each node with nonnegative, low-rank and sparse conditions. The LCLRR graph guarantees that the current node can be represented with a small number of other nodes. Meanwhile, it ensures the ability of capturing the global structure information as well as the local structure information of the images.In Section 4, we design a constrained image representation learning algorithm within variable ? based on semi-supervised NMF method. First of all, we obtain a new NMF model with label constraints by adopting the ?-divergence as the distance measurement. By setting different ?, the distance measurement in the proposed model is equivalent to several canonical distance measurements. Our method can be regarded as a generalized distance measurement over traditional approaches, thus it can better adapt to the complicated distributions in different visual datasets. Second, the optimization equation with parameter ? is solved by applying the Karush-Kuhn-Tucker conditions as well as the projected gradient method. The convergence of the algorithm is proved by theoretical analysis. Finally, we design a constrained algorithm that contains a variable parameter ? when assessing the performance of the algorithm in image classification. Extensive experiments show that our algorithm better adapts to different databases, and significant improvement on classification has been achieved by the proposed semi-supervised image representation learning model.