Image Saliency Detection Via Ranking With Local Linear Regression And Global Sequencing

In recent years, the image saliency detection has been paid attention and applied widely. According to the color, texture and other characteristics of the images, the image saliency detection generates a saliency gray map, in order to extract salient regions from the images. In the saliency gray map, the gray value of each pixel indicates the probability that the pixel belongs to the salient regions.This paper introduces a new transductive ranking method for image saliency detection, namely, ranking with Local Linear Regression and Global Sequencing (LLRGS). Saliency detection can be regarded as ranking super-pixels by some ranking mechanism. For every super-pixel, we make its color features basic data point for processing, and use a local linear regression model to predict the ranking value of its neighbors, making that adjacent super-pixel can have the similar ranking values. Besides, globally align local linear regression models from all the data points, and learn a Laplace matrix for image data ranking, thus, the saliency detection method basing on ranking with Local Regression and Global Sequencing is more effective and robust. With reference to prior maps, this paper uses a unified objective function to assign an optimal ranking value to each data point. At last, we get a saliency gray map.This paper experiment evaluates the introduced method on public datasets, namely, the ECSSD-1000 and MSRA datasets, which is contains some images with accurate human-labeled ground truth. The experimental results demonstrate that the introduced method can identify the salient objects from an image.