Salient Object Detection Via Multiple Instance Learning

Visual saliency detection aims to use computer and imaging device to simulate human vision system,which can quickly extract key information from plenty of images and provide efficient preprocessing for intelligent visual system.Although an enormous amount of salient object detection algorithm have been proposed,it is still challenging to develop an efficient and effective method for complex image.In this paper,a novel and unsupervised saliency model is proposed,which is based on object proposals and multiple instance learning(MIL).Object proposals are a series of candidate segments containing objects of interest,which are taken as preprocessing and widely applied in various vision tasks.However,most of existing saliency approaches only utilize the proposals to compute a location prior.We naturally take the proposals as the bags of instances of MIL,where the instances are the superpixels contained in the proposals.This method jointly considers the high-level object information of proposals and the mid-level cues of superpixels.Therefore,we formulate saliency detection problem as a MIL task(i.e.,predict the labels of instances using the classifier in the MIL framework).This method allows some flexibility in finding a decision boundary based on the bag-level representations and can identify salient superpixels from ambiguous proposals.To make saliency detection results more accurate,we integrate appearance similarity,structural smoothness and global uniqueness criteria into a single diffusion function that has a closed form solution.In addition,we introduce the MIL to an optimization mechanism,which obtains training bags from initial saliency maps produced by existing algorithms and iteratively updates training bags from easily distinguished samples to more difficult ones to learn a strong model.The significant improvement can be consistently achieved when applying the optimization model to existing saliency approaches.The proposed algorithms are evaluated on six popular saliency datasets.We first analyze each module of the proposed methods to verify their effectiveness.Then the proposed models are compared with the state-of-the-art saliency algorithms in terms of three evaluation metrics,which shows the proposed algorithms perform favorably.