Research On Ordinal Learning With Applications In Image Analysis

As a classical machine learning paradigm,ordinal learning(OL)has evoked increasing attentions in machine learning community,due to its wide range of applications.Although there have been numbers of OL methods proposed and obtaining good estimation results in various problems,the performance of such methods can be improved further due to that they have not sufficiently explored and utilized the prior knowledge involved in the problems and data(According to the well-known theorem of No Free Lunch: Only such learning algorithms are effective which are incorporated with the problem-specific prior knowledge).To this end,this dissertation concentrates on the OL,and attempts to embed the problem and data specific prior knowledge into modeling and thus promoting the learning performance.The main contributions of the dissertation are summarized as follows:1)Incorporating the spatial structure information into OL.Firstly,we make a systematic analysis on the works related to the use of spatial structure information of images,and classify them into three categories,i.e.,embedding in the distance metric,spatial structure regularizing,and directly operating on the images.Considering that the spatial structure prior information within images has not yet been utilized in existing OL methods which operate on the vectorized images(in the vectorizing transformation,the spatial structure of images is destroyed and thus the structure information is lost),we propose to construct OL models with incorporating the spatial information via the three strategies,by which the spatial information of images can be taken into account in OL.Finally,we experimentally compare the efficacy of the three strategies in image-oriented OL.2)Incorporating the inherent relations involved in ordinal cumulative attributes into OL.Through analyzing the ordinal cumulative attribute(CA)coding matrix of training data,we find that explicit and implicit structure relations exist within the coding matrix.To this end,we attempt to derive 0th and 1st relation matrices by performing difference computations along the adjacent rows(or columns)of the original CA coding matrix.Then,in order to incorporate the prior structure relations contained in the 0th and 1st relation matrices,we construct two corresponding kinds of regularization terms,coined as CA-oriented ordinal structure regularization(CAOSR)and CA-oriented adjacent difference orthogonal regularization(CAADOR),to mathematically depict the relation matrices,respectively.Next,we propose to construct OL regularized with the CAOSR and CAADOR.Finally,experimental results on ordinal human age estimation demonstrate the effectiveness of incorporating the inherent relations involved in the ordinal CAs into and consequently improving the learning performance.3)Constructing ordinal metric learning following the OL decision-making rule.Motivated by the OL decision-making rule,we propose to construct a novel ordinal margin metric learning(ORMML),in which the ordinal relationship between the data classes is well preserved and thus facilitates the subsequent OL decision-making.Then,to cope with more realistic scenarios where the data are sampled across distributions,we further extend the ORMML to its across-distribution counterpart(CD-ORMML).Finally,experimental results on several types of OL datasets demonstrate the superiority of the proposed methods over related state-of-the-art.4)Improving OL by means of using the auxiliary prior knowledge.In order to take advantage of the auxiliary prior information to improve the OL,we take the gender-aware age estimation as the research paradigm,in which the semantic discrepancy between human gender and aging semantic spaces as well as the gender difference between the male and female in their aging.Following such prior knowledge,we construct a unified gender-aware age estimation(GenAge)framework by using the gender-induced information as the auxiliary.Subsequently,in order to evaluate the efficacy of the proposed GenAge,we exemplify it and experimentally explore its effectiveness and superiority in age estimation over state-of-the-art methods.