Discriminative Structure Learning For Semantics Extraction From Images

With the explosion of digital data in Internet, it becomes particularly urgent to manage and leverage them in both effectiveness and efficiency. Pattern recognition is a promising way to accomplish it. However pattern recognition algorithms suffer from the existing limited label information. Thus how to perform discriminative learning effectively and efficiently from given priori is a great challenge for pattern recognition. This thesis focus on developing methodologies to bridge the ”semantic gap” between low-level visual features and high-level semantics and conducts discriminative learning for pattern recognition problems as follows:(1) Dictionary learning and collaborative representation based classification: Both dictionary and residuals play an important role in sparse representation based classification which is regarded as a special case of collaborative representation with sparsity constraints for pattern recognition problems. Inspired by the atom-level class separability which is guaranteed by sparse coding on the dictionary, we investigate the feature-level class separability on dictionary and residuals. Therefore we develop class specific feature-level component importance learning of dictionary and residuals by performing multi-objective optimization over sparse coding and improve discriminative ability of dictionary and residuals for classification problems.(2) Graph embedding and manifold learning based label inference: In manifold learning,graph has an ability of capturing data correlations and leads to discriminative learning along manifold data distribution. Thus we employ graph embedding to perform discriminative learning and bridge the ”semantic gap” between visual features and semantics for image classification. Specifically, we construct two different embedding formulations with graph Laplacian for data warping and label propagation, respectively. Data warping aims to learn discriminative features to capture global discriminative data structure and label propagation attempts to transfer semantic concepts of labeled samples to unlabeled ones for category inference.(3) Dictionary learning and manifold learning based scalable classification: Reconstruction based methods perform dictionary learning mostly in a supervised manner. However learning from limited label information cannot cover global requirements on priori. On the contrary, semi-supervised learning(SSL) strategy is capable to leverage a large amount of unlabeled data to approximate manifold distribution of data and perform learning from the captured manifold structure. Furthermore dictionary learning-based classification have a good property of scalability in test sample size which happens to be the very shortcoming of SSL. Therefore we propose bi-dictionary learning with smooth representation based label propagation with complementary of dictionary learning and SSL for scalable classification problems.(4) Manifold learning and kernel learning based classification: Kernel trick is a powerful technique to achieve non-linear correlations between samples in a complex data distribution.Thus we investigate discriminative learning with a combination of kernel learning and SSL.This approach employs strategies of adaptive constraints and seed propagation for achieving a discriminative kernel matrix and performing classification on the learned kernel matrix.Particularly, adaptive constraints helps in propagating supervisory information along the correlations between samples, while seed propagation implements full kernel matrix learning in block manner to guarantee the learning efficiency.(5) Applications on saliency estimation, image segmentation and retrieval: The target of computational intelligence is to facilitate human’s daily life. It is necessary to analyze user’s requirements and preferences in pattern recognition problems. We study to accomplish discriminative learning from user’s interactive inputs in image segmentation and image retrieval problems to meet individual user’s requirements. In interactive foreground extraction,users are allowed to provide markers on object of interest and its corresponding background.Then the markers are used as supervisory information to guide the learning and foreground extraction procedures. Similarly in interactive image retrieval, users are asked further to give relevance feedbacks on candidate images to direct learning for image ranking process.Furthermore, we apply the proposed kernel learning approach to saliency estimation. In saliency estimation, objectness completeness is guaranteed by saliency seeds propagation.Then objectness oriented saliency estimation is performed for salient region detection.