A crude to fine method to detect salient region

The task of salient region detection aims at establishing the most important and informative regions of an image. In this work, we propose a novel method that tackles such task as a process from superpixel-level locating to pixel-level refining. Firstly, we over-segment the image into superpixels and compute an affinity matrix to estimate the similarity between each two superpixels according to both color contrast and space distribution. The matrix is then applied to aggregate superpixels into several clusters by using affinity propagation. To measure the saliency of each cluster, three parameters are taken into account including color contrast, cluster compactness and proximity to the focus. We appoint the most salient one to three clusters as the crude salient region. For the refining step, we regard each selected superpixel as an influential center. Hence, the saliency value of a pixel is simultaneously determined by all the selected superpixels. Practically, several Gauss curves are constructed based on the selected superpixels. Pixel-wise saliency value is decided by the color distinction and spatial distance between one pixel and the curves' centers. We evaluate our algorithm on the publicly available dataset with human annotations, and experimental results show that our approach has competitive performance.