Salient Object Detection On Multiscale Learning And Sparse Coding

With the improvement of the technology and the development of society,computer vision has been permeating every aspect of people's life including the military,navigation,etc.The demands from the marketing has driven the innovation and development of this theoretical study constantly.Saliency detection is to simulate the process to the visual information of the brain.This dissertation aims in improving the precision of saliency detection universally.Having known current researches deeply,multi-scale random forest learning and weighted sparse representation are introduced to realize the saliency detection.First,random forest model is adopted to learn the feature information of train datasets.Meanwhile,to enhance the description of the outlines and details to the saliency target(s),the concept of the multi-scale learning is introduced to improve the performance of classifier learned.Through the test experiments,the multi-scale learning model do improve the results comparing with ones in single scale.However,the model needs optimization because of the unsatisfactory results of image evaluation indexes.Secondly,sparse coding is adopted as the foundation to achieve the optimization.The label results of the classifier to the image is adopted as the background dictionary.Aiming at reconstructing the different dictionaries,it is absolute to get optimum solution by iterative computation.And penalty function is used to construct the weighted sparse representation model,which is also constrain the dictionary reconstruction.Through,the test experiments,it is concluded that this comprehensive algorithm is capable of selfoptimization and the results on the test dataset is improved.in three different kinds of datasets,this algorithm performs excellently in saliency detection of most images.Comparing the other nine algorithms,the performance of this algorithm is better than one of those.However,it still exists the unsatisfactory detection results.According to analyzing the different experiments objectively,some specific possibilities are proposed to further study.Finally,there is a summary and outlook.