Fast DCT-based image saliency detection

Saliency detection is fundamental to many image processing and computer vision tasks, such as the applications of image re-targeting, retrieval and segmentation. However, highly accurate image saliency detection algorithms are also very computationally expensive, which limits their applicability for offline batch processing. In this paper, we employ the frequency domain algorithms to accelerate the saliency detection method with selecting fewer DCT coefficients. We combine the context-aware saliency detection approach with the efficient nearest neighborhood search strategy into a fast implementation. The runtime of the proposed approach is reduced from one minute to several seconds, and the created saliency map achieves the similar visual quality.