Research On Bottom-up Visual Saliency Detection Algorithms

Human visual system(HVS)has a remarkable ability to automatically and selectively pay more attention to important visual stimuli in natural complex scenes,due to the fact that it cannot fully process the tremendous amount of input visual information.In order to find out an analogous model mimicking human selective attention mechanism,researchers in physiology,psychology and computer vision have been making a good effort for a long time and proposed many computational models.However,the significant detection itself is a very complex issue,although the researchers have achieved a lot,achieve the intelligent processing of the human visual system must also have a more in-depth study.The major research content and contributions of this thesis are:Firstly,in view of the shortcomings of the previous model used in the linear weighted,combinning with existing evidences in neurophysiology,cognitive psychology and experimental conclusions of existing visual saliency models,this paper proposes significant detection method based on adaptive weights.The method using statistics principle,biology and multi-scale analysis method to predict the human attention points.And the proposed model of saliency at a pixel of interest is a data-dependent weighted average of dissimilarities between a center patch around that pixel and other patches,which overcomes the disadvantages of previous models using linear weighted.This paper proposed saliency model in predicting human eye fixations on common dataset and the comparison of ten state-of-the-art models has carried on the contrast experiment,experimental results demonstrate that the proposed method performs excellently for human eye fixation prediction.Secondly,through research on saliency detection via graph-based manifold ranking,the following conclusions were got:although the algorithm can get better results,they only consider the case of a single scale,which makes the algorithm has inherent limitations,mainly in inhibiting effect to the background,significant regional uniformity of fine.For this reason,this paper extends this method up to multi-scale,proposes significant detection based on multi-scale and manifold ranking.The method combines super-pixel technology,graph-based manifold ranking and multi-scale techniques to measure the saliency of images.The proposed saliency model is evaluate in predicting human eye fixations on common dataset and the comparison of eleven state-of-the-art models has carried on the contrast experiment,experimental results demonstrate that the proposed method performs excellently for salient object detection.