Saliency detection based on MI-KSVD

摘要

In this paper, we propose a visual saliency detection algorithm with MI-KSVD, a codebook learning algorithm that balances reconstruction error and mutual incoherence of the codebook. We first segment the images into superpixels by simple linear iterative clustering (SLIC), which can improve the efficiency and correctness of the progress. Then we calculate the reconstruction errors based on the initial background propagated from the boundaries of the image. We use a weighted sum of multi-scale region-level saliency as the pixel-level saliency in order to generate a more continuous and smooth result. Based on that, we further use object recognition as a vital prior to improve the performance of our method. Experimental results on three benchmark datasets show that the proposed method performed well to reach our expectations in terms of precision and recall.