Salient Objects Detection In Natural Images

The salient object detection in natural images is a fast developing research branch of computer vision community in a recent decade.Unlike traditional object detection,recognition and image segmentation,the goal of salient object detection is to detect and segment the salient objects in an image,instead of detecting and segmenting all objects or regions in an image.Moreover,in general,the output of salient object detection is also not the target bounding boxes of traditional object detection tasks,but the saliency map which indicates whether a region belongs to salient object.In essence,two fundamental key issues exist for salient object detection.First,how to compare the difference between different regions.Second,which features should be used to measure saliency,and how to define the mapping function between features and saliency value.Focusing on these two key issues,the following researches are carried out in this thesis.Firstly,in order to overcome the two shortcomings of many traditional algorithms that they often select contrast regions according to the “background prior” and using low-level features like colors,this thesis proposes a salient object detection method based on features spatial distribution(abbreviated as FSD).This method firstly computes the spatial distribution of high-level features with semantic information in an image,and then select the contrast regions for computing saliency.In contrast to the manner that relies “background prior” to select contrast regions,FSD can adaptively select contrast regions that cover more background regions,and is not sensitive to the spatial location of salient object in an image.In addition,FSD only uses high-level features that contain semantic information to detect saliency.As compared to many traditional algorithms that only use low-level features,the features adopted by FSD have more discriminative ability than low-level features.All these advantages help the FSD to effectively decrease the false and true postives.Secondly,in order to sovle the problem that the features used by traditional methods cannot effectively detect the saliency in complex images,this thesis proposes the salient object detection method based on multi-level features learning(abbreviated as MFL).It is based on the FSD,and additionally defines the multi-level features that include global contrast,local contrast and self-response.Thus,saliency detection can be carried out from perspective of both global and local contrast,as well as involving human’s experience knowledge.In addition,the mapping functions in MFL are learned automatically from samples,instead of being defined by human,which makes the algorithm can reasonably learn the manner to utilize the dimensions of features.These strengthes give the more powerful detection ability to MFL.Thirdly,this thesis also proposes the salient object detection method based on boosting object-level saliency(abbreviated as BOS),with respect to solving the problem that off-line learning method often don’t take pertinency on the training examples,and the procedures of both FSD and MFL are complex.BOS uses object proposals to replace superpixels as computing unit,which not only simplfy the process of generating saliency maps,but also improve the quality of saliency maps.More importantly,this thesis proposes the Boosting Forest algorithm to take pertinency on the training examples that are easy to be wrongly classified.By this manner,BOS can more effectively learn the mapping functions from complex images,and accordingly,BOS obtains the robustness to the hard negative examples.Finally,the speed of most deep learning methods are not fast,which due to their complex network sturctures,and their effecitiveness in detecting complex images is also low.In order to solve this problem,this thesis proposes the salient object detection method based on multilevel features generation network(abbreviated as MFGN).It is based on MFL,and uses a compact end-to-end network to realize the mind of extracting multi-level features in MFL.Moreover,its features have the ability to represent multi-scale context information,and the characteristic of simultaneously utilizing semantic and details information,all these give the MFGN the ability to run at a realtime speed of 33 FPS and effectively detect salient objects in complex image.