Salient Object Detection Via Extreme Learning Machine And Proposal Subset Optimization

Saliency detection is a preprocessing for many image processing tasks.It has been applied in image segmentation,visual tracking,and person re-identification and so on.In this paper,two algorithms are presented for saliency object detection,which integrate the low-level and high-level features and take advantage of both bottom-up and top-down models.The first algorithm proposed in this paper is a learning model based on extreme learning machine(ELM),which could be used for improving the performance of existing methods.Firstly,the input image is segmented into superpixels,each of which is represented by low-level features.Secondly,saliency maps generated by existing methods are taken as prior maps,where training samples including positive and negative samples are collected form multiply scales.These samples are used for training an ELM classifier,which could generate saliency map for an input image.Finally,multi-scale saliency maps are integrated to further improve the precision.Furthermore,an integration mechanism is put forward to take advantage of superiorities of multiply saliency maps by different methods.The second algorithm is based on subspace optimization of object proposals.Two modules,semantic attention(SeA)and spatial attention(SpA),are proposed to search for a subset of object candidate regions that have high probabilities to be the potential object.The SeA module focuses on the high-level semantic information and takes high-level semantic features from the Fast Region-based Convolutional Neural Network(Fast R-CNN)to select regions semantically similar with the foreground.While the SpA module pays attention to low-level spatial information.Regions of interest are detected by estimating their spatial similarities with salient objects predicted by contrast-based measurement.The final results are generated by integrating two maps from SeA and SpA for the purpose of incorporating spatial and semantic information for better performance.In order to demonstrate the effectiveness of the proposed algorithms,extensive experiments are conducted on five benchmark datasets.Compared with 24 state-of-the-art methods,the proposed algorithms have favorable performance in terms of three evaluation metrics,i.e.,the Precision-Recall curve,AUC and F-measure.