Detection And Fusion Of Visual Saliency Regions In Images

Saliency detection in neurobiology is a vehement research during the last few years.When dealing with the images,people perceive the images' information with different attention levels due to importance of the data usually known as region of interests(ROIs).Detection of visual saliency region is an effcctive way to realize ROIs in images.Saliency can be defined as "an attentional mechanism,which focuses on the worthiest part in the image." Several cognitive and interactive systems are designed to simulate saliency model.In spite of various existing state-of-the art algorithms for saliency detection,performance improvement in contrast to computational time cost and salient object segmentation in unconstrained and complex scenes remain chal-lenging.This dissertation focuses on different methods for the detection and fusion of visual saliency regions in images.Saliency detection has numerous applications in different image processing problems such as video and image compression,object recognition,image editing,image thumbnail creation,picture collage,image re-targeting and image retrieval.This disser-tation mainly relics on visual saliency detection,which aims to detect and segment significant prominent objects in bottom-up manner from natural images.In this dissertation,we have investigated three innovative bottom-up saliency detection methods to overcome the existing issues for saliency detection algorithms.First,we provide a General introduction and application of saliency detection algorithms in different image process-ing problems in chapter 1.An overview of existing bottom-up visual saliency detection methods is briefly discussed in chapter 2.Next,we propose two innovative bottom-up saliency detection algorithms for better estimation of salient objects and one novel algorithm for salient regions fusion in images.Experiments are conducted on different images datasets,our algorithms show better results as compared to the exiting state-of-the art saliency detection algorithms.In chapter 3,we propose a novel method(CFBF-SRD)that calculates the saliency by us-ing low level visual cues and Bayesian framework for better saliency maps detection based on foreground and background information.Bayesian framework leads to more accurate saliency estimation.In the proposed algorithm,color and luminance frequency features of RGB and CIE L*a*b*color spaces of the image are used to estimate the salient object and boundaries.Low-level visual features of image are employed and band pass filter is used to estimate and highlight the boundaries of salient region.Moreover,based on the likelihood probability,saliency is computed by applying Bayesian framework at each pixel by using foreground and background information.Extensive results confirm the reliability and efficiency of the proposed algorithm in terms of better precision_Recall,F-measure and computational time cost against state-of-the art saliency detection algorithms.In chapter 4,we propose a novel visual saliency detection method(SAMM-SRD)based on the Surroundedness and Absorption Markov Model.It provides foreground and background based saliency cues estimation by utilizing Absorption Markov model.In the proposed mod-el,saliency in images is estimated by utilizing the eye fixation salient features and absorption Markov chain.First the approximate area of salient object is predicted by surroundedness to the eye fixation points prediction.Secondly,Simple Linear Iterative Clustering algorithm(SLIC)is utilized to calculate superpixels and a 2-ring graph model is established on these superpix-els.Then,Absorption Markov chain is employed to estimate superpixels related to foreground and background regions.Finally,a guided filter is used to reduce the background noise and s-mooth the saliency map.The experimental results show that SAMM performed favorable against state-of-the art salient detection algorithms.Experimental results give significant evidence of better performance of the proposed model by showing the better precision_Recall,F-measure,MAE(minimum is better)Values.In chapter 5,we propose an salient regions fusion algorithm(DSET-SRF)by utilizing the DS-Evidence theory which is referred as "A synthetic fusion rule based on the DS-Evidence theory for saliency detection in images." It provides a synthetic way to unify the saliency maps with the most relevant saliency cues.In the proposed algorithm,different saliency maps are con-structed by using region-level,patches-level and pixels-level information.Then,DS-Evidence theory(evidence theory allows one to combine evidence from different sources and arrive at a degree of belief that takes into account all the available evidence)is utilized to fuse the salient pixels by using the foreground and background information based on calculated saliency maps.Experimental results demonstrate that the proposed fusion algorithm performs better as com-pared to the state-of-the-art algorithms.