Salient Object Detection:Computational Methods And Applications

Human visual system receives a mass of information from the outside world.As one of many remarkable abilities of human beings,selective attention allows us to extract important information from abundant visual inputs and interpret complex scenes in real time.Over the past several decades,many psychologists and physiologists have been trying to research how selective attention helps us to deal with a huge amount of visual information efficiently.The visual saliency was proposed and has been widely used in many computer vision tasks,such as object recognition,video summarization,image compression,image editing,and visual tracking,etc.As a branch of visual saliency computation,salient object detection is about detecting salient objects and segmenting the accurate regions of them in visual scenes.A good salient object detection method should meet high detection performance with low computational complexity.In-depth study of salient object detection can help to bridge the gap between low-level features and image contents,and make high-level scene understanding possible.This dissertation proposes two salient object detection methods and a co-salient object detection method,then applies saliency technique to small target detection in a color image.The main contributions of this dissertation are:(1)A regional principal color based salient object detection method is proposed.The proposed method first quantizes the input image using minimum variance quantiza-tion in RGB space,and computes global saliency by measuring color difference in LAB space.After segmenting the quantized image into regions using a superpixel segmenta-tion method,the proposed method chooses the most frequently occurring color of each region as the regional principal color to further compute local saliency.Finally,by ex-ploiting two kinds of spatial relationships,the proposed method can suppress non-salient regions meanwhile highlight central ones.The proposed method can reduce the compu-tational complexity of saliency computation and achieve higher performance with fewer quantitative colors.(2)A color name space based salient object detection method is proposed.The proposed method first converts the input image from RGB space to color name space,and obtains a saliency map by sequentially segmenting each color name channel and using a surroundedness cue based figure-ground segregation.Then the proposed method converts the input image to the color name image,from which two global color name features are derived and coupled with the surroundedness cue to generate a wighted saliency map.Finally,the proposed method combines the aforementioned two saliency maps using a truncation operation and a post-processing procedure.The proposed method can suppress background regions effectively and simultaneously highlight salient objects uniformly.(3)A hierarchical co-salient object detection method is proposed.The proposed method first constructs three image layers for each input image,and generates three single-layer saliency maps by measuring the saliency consistency between the saliency results of color names and background detection.Then these single-layer saliency maps are refined by invoking a global rarity cue of color names.The saliency consistency based combination and the color rarity based refinement are also used in the stage of multi-layer fusion.Finally,the proposed method exploits a color repeatedness cue to discard non-co-salient regions and generate co-saliency maps.In addition to being able to detect co-salient objects from a pair of input images,the proposed method can also detect salient objects in a single image.Experimental results show that the intra-and inter-saliency can benefit from the usage of color names.(4)A small target detection method combining regional stability and saliency is proposed.By exploiting five regional descriptors and four stability metrics,the proposed method uses a stability measure based regional stability algorithm to generate a set of small target candidates with their precise feature descriptions.Then a local color contrast based regional saliency algorithm is proposed,which can highlight small salient regions and simultaneously suppress background regions.Finally,the proposed method integrates the obtained stability and saliency maps,and uses the mean saliency of each small target candidate to remove false alarms.Experimental results show the validity of the proposed fusion mechanism.What's more,a new benchmark dataset for small target detection is labelled and published,which includes human annotated ground truth and experimental evaluation code.