Research On Visual Saliency Detection Methods And Its Applications

Visual saliency detection is a basic and complex technology in computer vision technology,which can guide computers to extract key information from image by simulating human visual habits,and provide the more efficient input for applications like image segmentation,image retrieval,image fusion and image editing.However,the study of visual saliency detection is an interdisciplinary subject which integrates cognitive psychology,neurology,mathematics,statistics,and so on.Although some research findings have been obtained,many problems still remain to be solved,such as uneven characteristic distribution,sample sensitivity,poor effectiveness and so on,which makes visual saliency detection still a challenging project.In order to solve these problems,three new saliency detection methods are proposed and applied into different image segmentation fields in this dissertation.The main contributions and innovations of this dissertation are as follows.(1)A saliency detection method based on dynamic guided filtering is proposed.Although the algorithms based on contrast filtering can detect the salient object with certain effectiveness,the accuracy of the detection decreases when the image characteristics are unevenly distributed.This problem can be solved by using the dynamic guided filtering.Therefore,a saliency detection method based on dynamic guided filtering is proposed.In order to tackle the problem of excessive number of iteration of the current dynamic guided filtering,a new and simpler kernel function is designed by combining the information of filtering results and input image.This could ensure a good structure transfer from input image to guided image,alleviate the detection error caused by uneven distribution of characteristics,and simplify the calculation model of dynamic guided filtering.Experimental results show that the proposed method is better in terms of detection accuracy and recall rate.Meanwhile,this method is efficient in saliency detection of remote sensing images.(2)A saliency detection method based on robust foreground seeds selection is proposed.The saliency detection methods based on manifold ranking are quite accurate and very efficient,but are also sensitive to the sample.Improper selection of sample may result in lower detection accuracy,or even detection failure.Therefore,a novel foreground seeds selection method is proposed,which used the intersection of convex hulls based on gradient and position of image regions to define the novel salient foreground regions,thus the foreground seeds are not obtained indirectly by the estimation of the background seeds.Experimental results indicate that the method is superior to the others in adapting to image complexity and in detection accuracy.(3)A saliency detection method based on improved LapSVM is proposed.In graph based methods,LapSVM is an advanced method for semi-supervised classification.However the adjacency matrix is usually fixed in the process of building a graph,which cannot well reflect the connection characteristics or similar features among nodes.In order to improve this,an improved LapSVM based on adaptive learning is proposed and used for saliency detection.The algorithm can find the accurate adjacency matrix and the labeled samples by iterative method,and the classification probability of the image region is used as its saliency value.Experiments show that this method has higher recall rate and is less time-consuming.(4)Three different image segmentation methods based on visual saliency are proposed.The first one is an aerial object segmentation method based on saliency and ICM.ICM is the simplified model of PCNN.When ICM is used for segmentation of image with low contrast such as aerial image,it is hard to focus on the object.In this paper,saliency factor is introduced into the design of ICM by improving the input item and dynamic threshold value,which is helpful in removing the unnecessary background interference in low contrast aerial images.Experimental results show that the proposed method stands out because it can acquire more accurate segmentation result.The second one is a natural object segmentation methods based on saliency and normalized cut.Normalized cut is a method commonly used for natural image segmentation.To handle the complexity and large data of natural images,the proposed algorithm can more accurately estimate the similarity between different superpixels by using the regional saliency to construct new weight value structure.Experimental results indicate that the proposed method is superior to others in terms of segmentation result and object extraction effect.The third one is a superpixel segmentation method based on saliency.Superpixels can provide an effective image representation using over segmentation,but regardless of the size of superpixel,the object of the image cannot be fully represented in a meaningful way,which makes it difficult to satisfy the needs of more advanced image analysis.Therefore,we improved the traditional superpixel method based on a more quickly saliency algorithm using improved LapSVM and merging strategy of superpixels.Experimental results indicate that the proposed method is superior to several superpixel segmentation methods in terms of image representation.