Saliency Detection Via Active Ranking In Subspace

Saliency detection task has witnessed a booming interest for years,due to the growth of the computer vision community.It has been applied to many computer vision task,such as image segmentation,image retrieval,image classification,image recognition and so on.Although much significant progress has been made,saliency object detection remains a challenging problem.There is a fact that a large number of labeled samples are required in the supervised learning training process,the labeling groundtruth is laborious.Meanwhile,there is redundant information in a large number of training samples,which will have a negative impact on the accuracy of the model.Active learning uses a selection mechanism to select samples with a large amount of information for training,achieving the goal of obtaining higher model accuracy using fewer training samples.Based on this,this paper combines kernelized subspace ranker with the idea of active learning,and proposes saliency detection via active ranking in subspace.This paper presents two pool-based active learning strategies,which consider the uncertainty and diversity of unlabeled samples respectively to select more informative samples to participate in training.By doing this,the number of training samples and the cost of labeling can be dramatically reduced.This paper proposes a method jointly using subspace mapping and Rank-Svm to learn a ranker by CNN features extracted from object-level proposals.During testing,it ranks the saliency of the image proposals,and weightedly fuses the proposals with the highest ranking scores.Last but not least,to refine the object boundaries,we introduce a superpixel-level post-processing method to further improve effectiveness.The proposed methodology reduces the cost of labeling,as well as the redendency of training set,which makes a obviously improvement comparing with our pre-proposed KSR method.The outstanding of this method comparing with some conventional methods is demonstrated by extensive experiments on four publicly available benchmark datasets.