Image Saliency Detection And Its Application In Image Retargeting

The salient targets of an image refer to objects or regions to which human eyes pay more attention instinctively. Saliency detection is the task of recognizing the salient targets, so the saliency detection algorithm should be able to simulate the human visual perception system and has the ability of intelligent selection and rapid processing. Image saliency detection has been widely used in image retargeting, image segmentation, image retrieval, object recognition and other aspects. The available algorithms have various drawbacks so it is of great interests to make deeper exploration of the saliency detection in the field of image processing and pattern recognition.Image retargeting aims at automatically adjusting the size of a given image to fit the arbitrary displays without introducing severe visual distortions. Seam carving algorithm is a popular content-aware image retargeting method. According to the saliency of the image it remove or copy the least salient seams to achieve the purposes of image retargeting. Therefore, saliency detection plays a vital role in seam carving.In the thesis, we first propose a new saliency detection method based on cartoon-texture decomposition of the image. It computes the saliency of an image by using the higher order statistics(HOS) of the cartoon component of the original image. The scale parameter ? of the cartoon- texture decomposition has a great influence on the final saliency. Therefore, we also give a method to adaptively select the scale parameter ? by using HOS.Then we apply the image saliency detection method to image retargeting. The HOS only detects salient edges. However, in order to protect the salient objects and their contexts, we propose a method to improve the saliency. Firstly, the obtained salient edges are used as a guide to generate a convex hull, which encloses the salient edges thus salient objects. So the convex hull can be used to detect the salient region of the image and is called saliency window. Secondly, a weight function is defined according to the saliency window. Finally, a new image importance map(IIM) is obtained by using the weighted function to modify the prior saliency. This IIM can highlight salient regions while suppress the saliency of background.Based on the newly defined IIM we present an improved seam carving algorithm. In order to illustrate the performance of our algorithm, we randomly selected 100 images from MSRA gallery, we evaluate our method by using both subjective and objective metrics, and the objective evaluation metrics is proposed by analysis of the change of the area and the color composition of the salient region. We also compare it with the-state-of-art methods. Experimental results show that our method can preserve the salient objects and edges better than the-state-of-art methods even if the background is heavily cluttered.