A Novel Fusion Framework for Salient Object Detection Based on Support Vector Regression

Abstract

Abstract Salient object detection is highly influenced with a variety of salient features extracted from an image. It is a challenging task in saliency detection to find the suitable contribution of each salient feature for salient object detection. In this paper, we propose a novel fusion framework for salient object detection based on support vector regression (SVR) that effectively combines salient features for computing saliency score. Firstly, we extract salient features using a state-of-the-art saliency method. Secondly, training samples are generated that include positive and negative samples. These samples are produced by applying two different thresholds on salient features and these thresholds are also obtained from salient features. Thirdly, the training samples are fed into support vector regression to learnt SVR model for obtaining a set of learnable parameters. SVR saliency score of the input image is computed by using learnt SVR model with salient features of the input image. Finally, SVR saliency score is refined by post processing operations to obtained enhanced saliency map. Experimental results are shown efficacy of the proposed framework against compared 10 state-of-the-art saliency methods on two publicly available salient object detection datasets viz. ASD and ECSSD. The proposed framework achieves 1.7% and 19.4% improvement of Recall and MAE score on ASD dataset and 11.1% and 3.2% improvement of Recall and F-measure score on ECSSD dataset from the compared highest performing method.KeywordsUnsupervisedMajority votingPost processingSaliency map