CoBRa: convex hull based random walks for salient object detection

Salient object detection is a challenging research area, which aims to highlight significant region of the visual scene more accurately and quickly. In this research direction, we propose a novel saliency detection model called Convex Hull Based Random Walks (CoBRa) approach. In the proposed model, an image is segmented into superpixels and Convex Hull is constructed based on the segmented image to roughly partition the segmented image into two regions: CH-foreground and CH-background regions and the centroid of the CH-foreground region is calculated. Then, initial saliency is computed by using two priors viz. contrast and center priors. Here, the proposed model exploits CH-foreground region centroid obtained by Convex Hull to computed center prior which is more efficient than image center. Afterwards, two thresholds are empirically obtained and applied on initial saliency map to produce two binary segmented images. Based on these two binary images, the proposed model collects foreground and background seeds. These seeds are further refined with CH-foreground and CH-background regions to produce reliable and effective seeds. Finally, a random walk is constructed with the determined seeds to generate a pixel-wise saliency map. The superiority of the proposed model is validated via extensive experimental results performed on six publicly available datasets viz. MSRA10K, DUT-OMRON, ECSSD, PASCAL-S, SED2, and THUR15K. The performance of the proposed model was compared with eight state-of-the-art methods in terms of Precision, Recall, F-Measure, Receiver Operating Characteristics (ROC), and Area under the curve (AUC). The proposed method outperforms or comparable with compared methods in terms of all the performance measures.

Automated summary

This paper proposes a Convex Hull Based Random Walks (CoBRa) approach for salient object detection. In this approach, an image is segmented into superpixels and a Convex Hull is constructed to partition the image into foreground and background regions. The centroid of the foreground region is calculated and used to compute initial saliency. Two thresholds are applied to produce binary segmented images, and foreground and background seeds are collected and refined. A random walk is then constructed to generate a pixel-wise saliency map. Experimental results on six datasets demonstrate the superiority of the proposed approach.