Salient Region Detection Based On Multi-view

Salient region detection is one of the important technologies in the field of computer vision.It is mainly used in the preprocessing process of other technologies,such as image retrieval,image classification,target detection,image video compression,image segmentation and so on.It is the pre-work of these technologies,the study of the salient region detection has a positive effect on the follow technologies.In recent years,salient region detection has been widely concerned and researched which has had much success,but the problem is still far from being resolved,there are three main problems: First,most of the algorithm only use the low-level features of image or one priori informationwhich limit the applicability of the algorithm;Second,there is still much room for improvement in the accuracy of the algorithm on complex datasets;Third,the algorithm runs slowly which results in a long time to process one image This thesis is devoted to solving the above problems and proposes a salient region detection algorithm based on multi-view.The main work and contributions are as follows:(1)Summarizes the existing algorithm,research status,and the general process,introduces the commonly used theoretical basis and related technologies.(2)Proposed a salient region detection algorithm based on multi-view which via the popular sorting algorithm(Manifold Ranking,MR)based on graph as the measure of saliency.Inspired by multi-view learning,we construct three views,include low-level contrast,background priori and object priori.In the view of background priori,introduce dynamic weighted fusion of four saliency maps.In the view of object priori,introduce the object detection and obtain the position information of the object in the image.After view combination,use saliency map ranking combined with the foreground rankingon initial saliency map for spatial optimization.(3)Optimize the complexity of the salient region detection algorithm.First,analy the time complexity of the algorithm and optimized at the algorithm level and the program level respectively.At the algorithm level,adjust the algorithm to design a multi-thread scheme.At the program level,there are many ways to optimize,including transfer float point data to fixed-point,replace the segment function,loop optimization and function optimization.(4)The algorithm proposed in this thesis is validated on six benchmark datasets.Compare with the state-of-the-art methods,the results of experiental on these six datasets demonstrate the proposed method performs well.Evaluated the effectiveness of the proposed improvement measures in this thesis,and verify the performance of the algorithm after speed optimization.