Water flow driven salient object detection at 180 fps. (April 2018)

• Record Type: Journal Article
• Authors: Huang, Xiaoming; Zhang, Yujin
• Is Part Of: Pattern recognition. Volume 76(2018:Apr.)
• Page Start: 95
• Page End: 107
• Publication Date: 2018-04
• Subjects: Salient object detection -- Minimum barrier distance -- Water flow -- Geodesic distance -- Saliency detection; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.10.027
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose one efficient approximate MBD computation method which shows higher performance in both speed and accuracy. Based on water flow driven MBD, a fast salient object detection method is proposed which achieve 180 fps speed. The performance of proposed saliency detection method outperforms other MBD based methods and state-of-the-art methods. Abstract: Minimum Barrier Distance (MBD) is one recently proposed saliency measure which provides more robust result than the geodesic distance. Due to accurate pixel-wise MBD computation needs high time expenditure, approximate implementation with raster scan and minimum spanning tree (MST) are proposed recently. Inspired by the natural phenomena of the water flow, we propose one efficient approximate method – water flow driven MBD. Seed pixels (such as image boundary in salient object detection) are assumed as source of water, the water can flow from source pixels to other pixels with different priority which determined by MBD cost, lower cost means flow earlier. MBD of each pixel can be computed during processing of water flow. Our MBD computation shows higher performance in terms both speed and accuracy. Proposed MBD computation only needs visit image once, while raster based MBD needs scan image three times, MST based MBD needs traversal on image twice and additionally needs time to construct a tree. Compared with two previous MBD approximation algorithm, our computation speed increased by 2.4 times and 5 times, while approximation error declined by 50% and 80%. Based on our fast MBD computation, a fast salient object detection method is also proposed. The accuracy of proposed method outperforms other MBD based methods, and shows better performance than the other state-of-the-art methods. The proposed method achieves 180 fps speed performance on four public datasets. In state-of-the-art methods, the highest speed performance is about 52 fps, our method shows 3.5 times speed improvement.