Research On Visual Saliency Detection And Its Application

Research found that human visual systems can quickly and efficiently find the interesting information from a cluttered natural environment.However,it is still a hot problem to simulate the human visual system to design visual saliency detection model in computer vision.In recent years,the visual saliency detection technology has a very important application value in object detection,object segmentation,image fusion,image compression and other fields.This thesis aims at researching images' visual saliency detection,proposing two saliency detection methods and applying the saliency detection into object segmentation.The major content and results of this thesis are as follows:(1)Based on the contrast characteristics of the image,we firstly propose a new saliency detection model in this paper.In this model,we utilize the conversion of the color space and discrete cosine transform to obtain the new opponent color features,and calculate the saliency of color features in the local contrast method.Secondly,in order to highlight the rarity of the whole image,the model also takes into account the global contrast difference,combing with the saliency of local contrast to more comprehensively reflect the low-level visual attention of the human eye.Furthermore,the algorithm also introduces the multi-scale operation,reducing the saliency of the background region.Experiments demonstrate that the proposed method can perform excellently for human eye fixation prediction and better than the existing algorithms.(2)The paper summarizes existing saliency models and considers the spatial characteristics,puts forward a model based on spatial distribution and spatial frequency.First,this thesis aims at researching the data distribution in space,using sparse-low-rank method to solve the sparseness,and taking an adaptive weighting algorithm to merge it.Then,the gaussian operator is adopted to filter image,combing with prior knowledge to gain the saliency of spatial frequency.And then,we integrate the saliency of spatial distribution with the saliency of spatial frequency to generate initial saliency map.To make the object of the saliency map can be more compact in space,the paper finally employs the graph regularization method to refine it.Experimental results show that our model can has a good quality.(3)Finally,we realize an automatic object segmentation algorithm based on visual saliency by using the result of saliency detection as heuristic information and applying to the interactive object segmentation system.The proposed algorithm can receive more accurate object segmentation results,while reducing the complexity of manual operation.