Font Size: a A A

Research On Ranking Model Of Image Understanding Based On Supervised Learning

Posted on:2016-08-17Degree:MasterType:Thesis
Country:ChinaCandidate:L L WuFull Text:PDF
GTID:2308330479451025Subject:Electronic Science and Technology
Abstract/Summary:PDF Full Text Request
Learning to rank has become research focus in the field of machine learning, natural language processing, multimedia and information retrieval. With the extensive application of machine learning, solving the ranking problem through machine learning has caused many scholars’ attention. Based on the domestic and international relevant research results, some new ideas of learning to rank are discussed. Besides, a new method is proposed for evaluating traffic congestion and facial attractiveness.Firstly, for solving the traffic congestion estimation problem, traffic scene congestion degree ranking calculation model is proposed based on supervised learning. Using supervised learning ideas, ranking functions per attribute(‘congestion’, ’average speed’) are learned. For traffic congestion degree ranking model, GIST feature of each frame training image is extracted, however, for average speed degree ranking model, firstly, video motion information is extracted and then GIST feature is extracted through frame differential method, finally, modified Ranking SVM projection model is introduced to learn a ranking function.Secondly, for the facial attractiveness automatic estimation problem, this paper proposes a personalized relative ranking model of facial beauty indicator. Firstly, a large number of volunteers give each training image a score to rank preliminarily; a generic model can be obtained. Then, a generic model is combined with users’ aesthetic to re-rank; a personalized ranking model based on users’ aesthetic can be achieved. GIST feature describes the global information of facial image and HOG feature shows the local information, the complementarily between them is considered. Therefore, compared with single feature type(Eigenfaces and Dense-SIFT), using a blend of GIST and HOG feature types can achieve a better ranking result.Thirdly, In order to improve the tolerance ability of local deformations for current feature extraction methods, we use local scattering convolution network(SCN) to extract feature for the prediction of facial attractiveness. Wavelet scattering transform, which can distinguish texture with the same Fourier power spectrum, is introduced to build non-linear translation invariants image representation by cascading wavelet transforms convolution with modulus and averaging operators. We build a relative ranking model of facial attractiveness to prove the effectiveness of this feature extraction method for predicting facial attractiveness.Finally, image ranking algorithm is studied in solving the problem of traffic congestion degree and the facial attractiveness automatic estimation in this paper. When solving traffic congestion problem, a supervised congestion degree ranking algorithm is proposed. Further, when solving the problem on facial attractiveness evaluation, this image ranking algorithm is improved and scattering convolution network is used for extracting facial features.
Keywords/Search Tags:learning to rank, GIST feature, traffic congestion degree, facial beauty indicator, personalized ranking model, blended feature, scattering convolution network
PDF Full Text Request
Related items