Saliency Detection Via Bootstrap Learning And Locality-constrained Linear Coding

Previous salient object detection methods used to only focus on low level features or use numbers of sample images along with their labeled ground truth to train a high level learning model. In this paper we propose two models where both low level features and learning model without using labeled ground truth are explored. The first one is a bootstrap learning algorithm in which both weak and strong models are exploited. Firstly, we construct a weak saliency map based on image priors to generate training samples for a strong model. Secondly, a strong classifier based on the samples exacted from multiple scales of an input image is learned to detect salient pixels. Then, results from multiscale saliency maps are integrated to further improve the detection performance. Lastly, the saliency maps of the weak and strong model are combined together as the final saliency map.The other algorithm that we propose for salient object detection is a novel coding-based measure by exploring both global and local cues for saliency computation. First, a bottom-up saliency map is constructed by considering global contrast information via low level features. Second, utilizing a locality-constrained linear coding method, we formulate a top-down saliency map by computing reconstruction errors. Finally, the above two maps are combined to better exploit the local and global information.The experimental results on six benchmark datasets show that the proposed methods per-form favorably against 19 state-of-the-art saliency methods in terms of three popular evaluation measures, i.e., the Precision and Recall curve, Area Under ROC Curve and F-measure value. Furthermore, we show that the two proposed methods can be easily applied to other state-of-the-art saliency models for significant improvement.