Research On Salient Region Detection Methods

In the computer vision community,saliency detection refers to modeling the selective mechanism of human visual attention.The outputs of a saliency detection algorithm are usually termed as saliency maps,which represent the conspicuousness levels of different areas in a scene.Since saliency detection is an effective way to acquire potential regions of interest that may attract human eyes,its numerous applications range from object detection and recognition,image compression,video summarization,to content-based image editing and image retrieval.This thesis focuses on one research branch in saliency detection community called salient region/object detection,whose aim is to detect and emphasize holistic salient objects in the obtained saliency maps from input natural images.Despite that nowadays many models and algorithms on salient region detection exist and a consistent rapid progress in this field has been made over the past decade,improving the detection performance in complex and unconstrained scenarios still remains challenging.This thesis proposes five novel methods towards salient region detection,each of which is driven by specific motivations that try to solve certain shortcomings observed from existing models.Meanwhile,several new theories have been applied to salient region detection.The main contributions of this thesis are summarized as follows:1.A method that utilizes the characters of color contrast and color distribution is proposed.The proposed superpixel-based computational scheme effectively computes the color contrast and distribution attributes in a unified manner,which leads to complementary performance and hence improved detection accuracy.2.A novel method is proposed by employing geodesic distance to maintain the local connectivity of objects.Therefore,visually coherent saliency maps are achieved.Given an initial superixel-based coarse saliency map,the saliency of a superpixels is propagated to all superpixels in the image,where the extent of propagation between two superpixels is manipulated by geodesic distance.The proposed propagation technique,called geodesic propagation,is related to geodesic filtering.It is able to recover oversuppressed saliency parts of an object meanwhile suppress falsely detected background in the coarse map.As a result,enhanced salient region detection is obtained.3.Towards better grouping of objects and background,a method based on Normalized graph cut(Ncut)is proposed for saliency detection.Since the Ncut partitions a graph in a normalized energy minimization fashion,resulting eigenvectors of the Ncut contain good cluster information that may group visual contents.Motivated by this,the proposed method directly induce saliency maps via eigenvectors of the Ncut,contributing to accurate saliency estimation of visual clusters.This is different from existing salient object detection models that usually favor over-segmented regions upon which saliency is computed.The proposed method implements the Ncut on a graph derived from a moderate number of superpixels.This graph captures both intrinsic color and edge information of image data.Starting from the superpixels,an adaptive multilevel region merging scheme is employed to seek such cluster information from Ncut eigenvectors.With developed saliency measures for each merged region,encouraging detection performance is obtained after across-level integration.4.To improve the diffusion performance,a robust diffusion scheme is proposed,referred to as manifold-preserving diffusion(MPD).MPD is built jointly on two assumptions for preserving the manifold used in saliency detection.The smoothness assumption reflects the conditional random field(CRF)property and the related penalty term enforces similar saliency on similar graph neighbors.The penalty term related to the local reconstruction assumption enforces a local linear mapping from the feature space to saliency values.Graph edge weights in the above two penalties in the proposed MPD method are determined adaptively by minimizing local reconstruction errors in feature space.This enables a better adaption of diffusion on different images.By utilizing MPD,a two-stage saliency detection scheme is further introduced,referred to as manifold-preserving diffusion-based saliency(MPDS),where boundary prior,Harris convex hull,and foci convex hull are employed for deriving initial seeds and a coarse map for MPD.5.Observing that saliency detection can be treated as a continuous labeling problem,a novel data-driven scheme based on a special conditional random field(CRF)framework called continuous CRF(C-CRF)is proposed,where parameters for both unary and pairwise potentials are jointly learned.Therefore,the proposed C-CRF learning provides an optimal way to integrate various unary saliency features with pairwise cues.This differs from existing methods which employ manually designed parameters,or learn parameters partly for the unary potentials of the CRF.Additionally,a novel formulation of pairwise potentials that enables learning weights for different spatial ranges on a superpixel graph is investigated.