Saliency Detection in Textured Images

Recently, salient object detection has achieved significant development. Unfortunately, existing methods mainly depend on color differences, not effective for textured images. This is because the visual patterns of textures cannot be well measured with existing methods. In this paper, we address this challenge by using windowed inherent variation to capture texture information and meanwhile performing edge-ware superpixel segmentation. Thus, superpixels can be well utilized to distinguish contents from textures for improving saliency detection. We further employ background and foreground priors via graph-based manifold ranking to improve saliency estimation. For evaluating our method, we collected 200 textured images from literature to build a dataset. With both qualitative and quantitative evaluations on our dataset and other two benchmarks, the results show that our approach can significantly promote saliency detection in textured images, compared with the other state-of-the-art methods.