Research On Salient Object Detection

While people are understanding images or videos,their visual attention systems are able to quickly locate the most eye-catching objects,which are often named salient objects.In recent years,automatic detection and segmentation of salient objects from images or videos have been a hot research subject in the field of computer vision.In this thesis,the following researches on salient object detection are carried out: ultra-contrast based salient object detection algorithm,scene-dependent saliency object detection algorithm,object boundary guided saliency refinement,and saliency quality assessment based on deep convolutional regression network.The specific research contents and innovations include the following aspects:1.The traditional contrast-based salient object detection algorithms focus on highlighting the most dissimilar regions and generally fail to detect complex salient objects.We propose a salient object detection principle from existing contrast-based methods:dissimilarity produces contrast,while contrast leads to saliency.Guided by this principle,we propose a generalized framework to detect complex salient objects.The proposed framework is capable of flexibly integrating different kinds of region dissimilarity definitions,region contexts,and contrast transformations.2.Although there exist large variations in the category,location,and scale of salient objects,most supervised algorithms train a single holistic saliency detector and apply it to all test images.These saliency detectors are prone to missing those salient objects that seldom appear in the training set,as minimizing the loss function of a holistic saliency detector will result in the importance of the minority training images being neglected.We propose that specialized saliency detectors should be utilized for each test image.Specifically,considering that the category,location,and scale of salient objects dominate the appearance and structure of an image,we propose a scene-dependent saliency detection framework,aiming at detecting a salient object based on its corresponding scene.3.To improve the quality of initial saliency maps,existing algorithms typically refine them via a neighbor-constrained smoothing model,which assigns similar saliency values to neighboring regions.Since the adjacent regions could also cross the boundary between the salient object and background,these spatial distance-based methods easily cause false detection by involving the background regions that are close to the salient objects.To address this problem,we propose a boundary-guided optimization framework to jointly improve the region smoothness and correct the false detect regions.4.High-quality saliency map plays an important role in boosting many other computer vision tasks,such as object detection and segmentation.To assess a saliency map's quality,the only way is to utilize a full reference metric,i.e.,compute it with the groundtruth reference map.However,in the real-world applications,the ground-truth reference map for the saliency region is unavailable,which brings urgent demands for developing no reference saliency quality metric.We propose a deep saliency quality assessment network(DSQAN)to predict the saliency quality scores directly from saliency maps.As a direct application of the proposed DSQAN,the predicted saliency quality scores are first utilized to choose the optimal saliency map from a set of saliency map candidates.The experimental results on the MSRA10 K data set demonstrate that our proposed method could precisely predict the saliency quality.