Image Saliency Detection Theory And Applications

Saliency detection aims to quickly identifies subsets within the scene that contains significant information, and paves the way for its relative applications. In recent years,with the development of computer vision, as a branch saliency detection plays an important role. Saliency detection has been widely used for image retargeting, image cropping, image compressing and target recognition, meanwhile the extensive applications further improve the theory of it.Sparse representation based saliency detection plays an important role in salient object detection, it assumes that the background regions of an image can be sparsely represented by the background dictionary with small representation errors but that the salient regions cannot. Upon this, using features of the backgrounds as dictionary and sparse representation error as saliency measure, we get the saliency map of an image.One of the most valuable application of saliency detection is seam carving based image resizing. Seam carving chooses seams with less saliency to be carved according to the saliency map and get the resized image.This thesis studies the specific demands that seam carving needs on saliency detection,then gives two different ways of saliency detection to generate the demanded saliency map using sparse representation and wavelet decomposition.Different from the single gaussian prior saliency distribution, We assume that the prior saliency distribution of an image is multi-gaussian. We take the regions with less prior saliency as backgrounds, using color features of the backgrounds as dictionary to sparsely represent the whole image. the representation errors are used as saliency measure, meanwhile together with a specialized bevel edge energy map, we generate the tailored saliency map.We also present a salient profile detection method, which describes the salient profile of an image using wavelet decomposition, local and nonlocal feature descriptor and generate the salient profile map. In seam carving salient profile map protects the salient objects of an image from being carved,thus the resized image looks natural.Using above saliency maps for seam carving, Experimental results show that these two new methods are successful incorporation of saliency detection and seam carving, and gives good solution to the existing problems of seam carving such as: distortion and mismatch.