Image Saliency Detection Based On Synchronized Graph Ranking And Multi-task Learning

With the advent of the big data,the network society produces a large amount of data every day.How to deal with the increasing image data quickly and effectively has become an urgent problem to be solved.It is worth mentioning that the main task of image saliency detection is to make the computer locate the salient regions of the image by simulating human visual attention mechanism,and filter the non-salient region,so that only the region of human interest in the image is subjected to subsequent processing.Therefore,image saliency detection can effectively reduce the complexity of image data and improve it processing efficiency.As a basic research topic in the field of computer vision,image saliency detection attracts a large number of researchers.Since the problem of image saliency detection has been raised,researchers have proposed many algorithms with excellent performance.In particular,due to the successful application of deep learning method in the field of computer vision in recent years,image saliency detection has opened up a new way of research.More and more deep learning methods have been applied to image saliency detection,and have achieved breakthrough research results.At the same time,because the image saliency detection algorithm is simple and effective,it can easily be combined with the research in other fields of computer vision,so it can assist algorithms in other fields to achieve better results.For example,the image saliency detection method is used to realize the initial object location in the object detection and recognition tasks,thus improving the accuracy of the algorithm,and reducing the complexity of the algorithm.Therefore,the research on image saliency detection method can not only promote the improvement of complex image detection results,but also provide technical support for processing massive image data.However,the existing image saliency detection algorithms have limitations in the complex environment.When the image has a complex background,or the foreground is similar to the background in the image,the current image saliency detection algorithm does not obtain the best result.In view of the above problems,this thesis proposes two effective image saliency detection algorithms:The first one is an image saliency detection algorithm based on the synchronized graph ranking model.Aiming at the deficiency of the existing saliency detection model based on manifold ranking,namely the error detection problem when the contrast between foreground and background is small and the background is complex,it starts from the high contrast between the salient object and the background in the image,co-optimizes the background salient value and the foreground salient value,uses the background information to highlight the foreground information,and proposes an image saliency detection algorithm based on the synchronized graph ranking model,so as to obtain more accurate detection results.Experiments on five open data sets show the effectiveness of the proposed algorithm.The second one is an image saliency detection algorithm based on multi-task synchronized graph ranking model.In the process of image saliency detection,a variety of feature fusion methods are generally adopted to make features more discriminant.Traditional feature fusion directly processes multiple features in series,which is unable to make full use of features and easy to cause feature redundancy.In order to effectively realize the complementarity of each feature,this thesis proposes a multi-task synchronized graph ranking model and a new iterative optimization algorithm to solve the model.Through experimental analysis,it can be found that the traditional features tend to describe the low-level features of the image,which is beneficial to obtain the details of the image.The deep feature is focused on describing the semantic information of the image,which in favor of the accurate positioning of the object.Therefore,we choose traditional features and deep features as optimization objects.The specific steps are as follows:first,the composition is carried out on different feature levels;Then,the multi-task synchronized graph ranking model is used to dynamically update the weights of the images composed at different feature levels,so as to effectively combine the advantages of different features;Finally,the effective values of the model iterative convergence are used to determine the salient region of the image.Experiments on five open data sets show the effectiveness of the proposed algorithm.