Sparse Representation, Low Rank And Their Applications On Image Understanding

Spare representation and low rank representation are recently two hot topics in the field of image understanding, which imitates the spare strategy in human visual system and also can reveal real structure of data behind the observation. In this dissertation, we focus on several drawbacks of their applications on image understanding and propose methods of solving the problems which include:1) improvements on Bag-of-Words (BoW) model. The codebook in this model is usually noisy because many noises are involved when it’s constructed,; and is not discriminative for classification beacause it’s generated on data from all classes. By considering the effect of feature selection in sparse representation, we choose it as a way to select valuable words in the codebook. Additionally, we also learn a weight for each class with this technique. Experiments on several widely used datasets demonstrate superior performance of the proposed approach over standard BoW method.2) propose of new kind of sparse representation with the class constraints on coefficients. Traditional sparse representation is based on whole basis, which clearly is less of information of the source of basis. To take the advantage of this information, we add class constraints on coefficients which penalize the non-zero coefficients with different classes from the original sample’s class. The problem is convex and is solved by common coordinate descent method. Finally we verify the effectiveness of proposed algorithm by experiments on three standard databases.3) research on attribute-based image classification. Visual attributes have recently been proposed as a new perspective of understanding objects. In view that attributes of an object are always sparse and occasionally noisy, we introduce an t2,1-Norm regularizer into conventional linear model to weaken the influence of noisy attributes and be robust to outliers. The problem is sovled by using simple Lagrange Multiplier method. Morover, the attribute prediction is cast into multi-label learning problem since each image usually has labels of more than one attribute. Extensive experiments on three standard image databases with attributes demonstrate the superiority of proposed approach.4) robust object co-detection. Object co-detection aims at simultaneous detection of objects of the same category from a pool of related images by exploiting consistent visual patterns present in candidate objects in the images. The related image set may contain a mixture of annotated objects and candidate objects generated by automatic detectors. Co-detection differs from the conventional object detection paradigm in which detection over each test image is determined one-by-one independently without taking advantage of common patterns in the data pool. We propose a novel, robust approach to dramatically enhance co-detection by extracting a shared low-rank representation of the object instances in multiple feature spaces. The representation is based on a linear reconstruction over the entire data set and the low-rank approach enables effective removal of noisy and outlier samples. The extracted low-rank representation can be used to detect the target objects by spectral clustering. Extensive experiments over diverse benchmark datasets demonstrate consistent and significant performance gains of the proposed method over the state-of-the-art object co-detection method and the generic object detection methods without co-detection formulations.