Abstract: Aiming at the problem that the traditional Nonlocal Mean (NL-Means) filtering algorithm lacks the consideration of human vision system,this paper proposes a saliency region detection algorithm based on the improved Nonlocal Mean filtering algorithm.Firstly,the primary graph is generated by using the color uniqueness as the low-level clue.Then,the objectness estimation is utilized to generalize a set of object proposals,and the object measure estimation score is pooled to each superpixel patch to generate the foreground and background saliency map as the high-level clues.Finally,the improved Nonlocal Mean filtering framework with the object-level cue which multi-scale fuses the low-level saliency map with high-level saliency map generates the final saliency maps.A series of experiments are done on MSRA-1000,ECSSD dataset and results show that,compared with the current popular detection algorithms,the proposed algorithm not only suppresses the background areas while uniformly highlighting the salient objects,but also improves the performance evaluation of precision,recall and F-measure values.