U-FIN: Unsupervised Feature Integration Approach for Salient Object Detection

Salient object detection is a challenging research field in computer vision. The existing saliency detection methods generally focus on finding feature maps for saliency computation. However, the combination of these feature maps significantly improves salient region(s) detection. In this paper, we propose a novel feature integration approach called U-FIN in which final saliency map is obtained by a weighted combination of individual feature maps. The proposed approach works in three phases viz. (i) artifact reference (AR) map generation (ii) weight learning and (iii) final saliency map computation. Firstly, AR map is produced using majority voting on the individual feature maps extracted from the input image. Secondly, linear regression is employed for weight learning which is used in the next phase. Finally, the individual feature maps are linearly combined using weights learned in the second phase to generate the final saliency map. Extensive experiments are conducted on two benchmark datasets, i.e., ASD and ECSSD to validate the proposed feature integration approach. The performance is measured in terms of precision, recall, receiver operating characteristic (ROC) curve, F-measure and area under the curve (AUC). Extensive experiments demonstrate the superiority of the proposed U-FIN approach against nine state-of-the-art saliency methods on ASD dataset and comparable on ECSSD dataset with the best performing methods.