Visual Saliency Research For Natural Image

The problem of saliency detection for natural images has a long history.The concept of visual significance detection,which is not clearly defined in the field of scientific research,is generally considered to be based on the working principle that computer vision technology can effectively simulate the human visual system,to locate effectively and identify interesting regions or targets in natural images.This process is defined as visual saliency detection(or saliency detection).The biggest difference from the industrial image is that the background and the significant regions in the natural image cannot be obtained by prior distribution,and this makes the saliency detection process more difficult.The purpose of this paper is to efficiently detect saliency regions in natural images and to simulate human vision to some extent.For some basic applications,such as remote sensing location,target tracking and medical imaging,it is of great significance.In recent years,new relevant methods of saliency detection emerge in a large number.As for the research thought,it can be divided into two parts: 1.lateral analysis thought.It takes advantage of the characteristics of the background and the target in the image to make a difference.It can also be differentiated through combining the edge,texture,color and other effective information,so as to achieve the purpose of saliency detection.2.longitudinal analysis thought.It uses the depth information between the relevant image groups and figures out common saliency objects,so as to complete the task of saliency detection.The research methods can be divided into the following categories: 1.analysis method based on frequency domain transformation,2.analysis method based on graph,3.analysis method based on machine learning,etc.Some of the existing models can only solve some natural images with relatively fixed background and target.When it comes to the multi-target detection under complex background,the information contained in the background can not be effectively described,or the similarity between the target information and the background information is very high.All these will lead to the failure of the saliency detection.In this paper,the saliency propagation algorithm is proposed to solve saliency detection problem under complex background.Secondly,for RGB image detection,the co-saliency detection model does not take into account the depth information hidden in the image.Based on the full consideration of depth information,an iterative algorithm framework is proposed to solve the co-saliency detection problem.The research achievements and contributions of this paper are as follows:1)The saliency detection model based on the two-dimensional fractional order Fourier is proposed to make up for the defection problems exposed by the frequency domain method.(The frequency domain method is relatively effective when the background color change is single and the texture is clear.)The concept of noise sensitivity NSS is defined to characterize the robustness of the algorithm.And the effectiveness of the method is verified by experiments.2)Aiming at the problem of saliency detection under complex background,this paper proposes a discriminative saliency propagation algorithm(DSP).By creating a similarity measurement in the new feature space,the similarity metric between the background and the target is depicted,and the rough saliency map based on the background prior is further generated.The saliency propagation mechanism is introduced and the final saliency map is obtained by further refinement.It is verified by analysis that the framework described in this paper is also applicable to other existing salient models and can improve the performance of these models.At the same time,this method has better performance when evaluated by different indexes.3)In view of the co-saliency detection problem of RGBD images,this paper proposes an iterative RGBD co-saliency framework,which can transform the existing two-dimensional saliency model into a RGBD co-saliency form.The framework covers three processing schemes: addition scheme,deletion scheme and iteration scheme.The addition scheme is used to optimize the single saliency map,and a new descriptor is used to introduce DSP depth information into the framework.The aim of the deletion scheme is to use a common probability function to capture the constraints between images and to suppress non-public areas.The probability function is described as the possibility that each superpixel image belongs to the public area.Finally,an iterative scheme is designed to obtain a more uniform and consistent co-saliency graph.Through the comprehensive comparison and discussion of two groups of RGBD data sets,it is proved that the method is superior to other advanced saliency models and co-saliency models.4)An improved co-saliency detection framework is proposed to find common areas in a set of Images.The framework captures advanced semantic information by using Depth Co-saliency Network to transfer some significant prior knowledge to achieve uniform characterization of common objects and their corresponding boundaries.