Study On Orientation Contrast Based Boundary Detection And Image Classification

In this dissertation, we focus on two important problems in natural scene under-standing:image boundary detection and image classification. Usually employed as a preprocessing step for high-level vision tasks, boundary detection algorithms output com-pressed representation for images. While in image classification, an image is classified to some predefined class using the extracted features. Although the two image processing tasks seem to be weakly related, we argue that feature representation is the key factor that links the above two tasks. After analyzing the advantages and disadvantages of the existing algorithms, we propose effective algorithms for image boundary detection and image classification from the feature perspective. Our main contributions are:(1) The boundary detection task has been extensively studied in the field of computer vision and pattern recognition. Recently, researchers have formulated this task as supervised or unsupervised learning problems to leverage machine learning meth-ods to improve detection accuracy. However, texture suppression, which is impor-tant for boundary detection, is not incorporated in this framework. To address this limitation, and also motivated by psychophysical and neurophysiological findings, we propose an orientation contrast model for boundary detection, which combines machine learning technique and texture suppression in a unified framework. Thus, the model is especially suited for detecting object boundaries surrounded by natural textures. Extensive experiments on several benchmarks demonstrate the improved boundary detection performance of the model. Specifically, its detection accuracy was improved by10%on the Rug dataset compared with state-of-the-art unsuper-vised boundary detection algorithm. And its performance is also better or at least comparable with previous supervised boundary detection algorithms on BSDS500dataset.(2) Recently, sparse coding-based algorithms have achieved high performance on sev-eral popular image classification benchmarks. However, few works have used the spatial information of local image descriptions to improve either coding or the pool-ing operation. This paper proposes a novel spatially constrained coding scheme, which employs the m-nearest neighbors of a local feature in the image space to im-prove the consistency of coding discriminant ability. Specifically, with this coding strategy, similar image features will be encoded with similar visual words, which reduced the stochasticity of conventional coding strategy. Extensive experiments on the UIUC sport event,15natural scenes and the Caltech101database using sev-eral popular algorithms suggests that the image classification performance can be ubiquitously improved by incorporating the proposed spatially constrained coding scheme, firmly suggesting the generality and usefulness of the proposed approach on image classification.(3) We study the image feature representation problem in sparse coding algorithm which is how to select the optimal pooling regions for the codewords. We show that the Viola-Jones algorithm, which is well-known in face detection, can be tai-lored to learning pooling regions for the sparse coding algorithms. Specifically, using the boosting approach to learning pooling regions, image classification per-formance can be ubiquitously enhanced on several benchmarks (UIUC sport event,15natural scenes and the Caltech101dataset) to the state-of-the-art, using only low dimensional features and small codebook sizes. Furthermore, the "salient pooling regions" can be obtained explicitly.(4) Recently, mid-level features extracted from discriminant patches have been demon-strated powerful in image classification problems. We combine local features and mid-level features to obtain a more powerful feature representation for image clas-sification. Specifically, we first segment an image into several coherent regions by SLIC superpixel method. Then, in each region, we learn a subspace from all the local features falling in that region. The basis vectors spanning that subspace are combined to generate a new mid-level feature by the subspace-to-point mapping algorithm. To merge the two types of features, we construct two dictionaries:one for local features and the other for mid-level features, and the two types of features are concatenated to form the final image representation. Extensive experiments on the Caltech101and the Caltech256databases using two popular algorithms sug-gest that by combining the two types of features, the image classification accuracy can be significantly improved, which demonstrates the usefulness of the proposed method for image classification.