Salient Object Detection Based On Unsupervised Learning

The attention mechanism in Human Vision System(HSV)has gained increasing focus recently,one important branch among which is saliency detection.By detecting salient regions or objects in a scenario,researchers are able to study such an attention mechanism.Moreover,detected salient regions contain the essential information of the scene,which makes it easier for machines to understand the physical world and benefits higher-level tasks in computer vision.Existing unsupervised approaches for salient object detection starts from capturing features for visual inputs such as images,which include contrast information,boundary connectivity and so on.Generally,they can be very efficient and do not require labelled data,but the detected salient objects are always not complete enough.Besides,these methods often fail to obtain satisfactory detection results under cluttered background due to their heavy dependence on single information source or prior.To address these issues,in this paper,a compactness diffusion model is proposed.It detects salient foreground regions and redundant background regions separately,and then combines the results.By constructing a global graph,a statistic measure named compactness is introduced to detect those compact regions within an image.By constructing local graphs,the compact regions are set as seeds and diffused by a quadratic energy model,which generates continuous and uniform saliency map.This way,it avoids detecting incomplete results and is more robust to cluttered background.Experiments conducted on three public datasets validate the efficacy and effectiveness of the proposed model.Though the compactness diffusion model can overcome the aforementioned issues,it is difficult to handle the cases where two or more salient objects exist.Therefore,a hierarchical salient detection model based on low-rank matrix decomposition is further proposed in this paper.The model consists of a coarse module and a refinement module.The coarse module is devised as a low-rank decomposition model with Laplacian regularization,while the refinement module learns a projection matrix based on the generated coarse saliency map and also the spatial relationship among image regions.In experiments,the coarse module is compared with other approaches that are based on lowrank matrix decomposition.Combined with the refinement module,the proposed model is systematically compared with the other twelve state-of-the-art methods on three public datasets.The results demonstrate the efficacy of the proposed model,especially under the scenarios of multiple salient objects.