Human Image Data Augmentation Based On And-Or Graph

Human image data is an important kind of data,which has the characteristics of rich diversity and complex label.In practical application,it has become a trend to expand human image data by image generation.However,the current human image data enhancement algorithms can not meet the needs of practical tasks,because some of them lack the appearance diversity of the extended image data,some of them do not consider the posture diversity of the extended image data,and some of them are used in harsh conditions.In order to solve the problem above,this paper introduces the concept of and-or graph(AOG),and studies the application of and-or graph in the field of human image data enhancement.The research of this paper mainly includes the following aspects:(1)In this paper,a new human image data enhancement algorithm based on AOG is proposed.The algorithm uses AOG to construct human image representation model,and expands human image data by sampling.In the construction of AOG,the human image is divided into two parts: appearance and posture.A human appearance AOG is used to represent the appearance diversity of human image,and a human skeleton AOG is used to represent the posture diversity of human image.Thanks to the combined reconfiguration property of AOG,the human appearance AOG and human skeleton AOG can learn from limited data and generate more human appearance samples and three-dimensional human posture samples.After that,based on the constructed human appearance AOG and human skeleton AOG,a new human image is generated by combining the human appearance samples and the threedimensional human posture samples,so as to realize human image data enhancement.(2)The human appearance AOG was constructed.In the existing AOG based human representation model,phrase structure grammar and attribute grammar are used to describe human appearance,but the existing phrase structure grammar and attribute grammar are not suitable for human image data enhancement.By improving the phrase structure grammar and attribute grammar in the existing literature,this paper constructs the human appearance AOG.The human appearance AOG uses pixel block without background as image template,and adopts the decomposition structure referring to the human motion model,which is more suitable for human image data enhancement.In order to ensure the convenience of the human appearance AOG,a matching learning method is proposed,which allows learning from ordinary human image data.In addition,this paper also presents a sampling method with attribute constraints,which can generate human appearance samples with different attribute constraint levels.(3)The human skeleton AOG was constructed.In the existing AOG based human representation model,dependency grammar is used to describe human posture,but the existing dependency grammar is not suitable for human image data enhancement.By improving the dependency grammar in the existing literature,this paper constructs the human skeleton AOG.The human skeleton AOG describes the three-dimensional human pose sample space,and reasonably expresses the effective range and probability distribution of the three-dimensional human pose,which is more suitable for human image data enhancement.In order to ensure the convenience of human skeleton AOG,a matching learning method is proposed,which allows learning from two-dimensional human posture dataset.In addition,this paper also presents a corresponding sampling method,which is used to generate three-dimensional human posture samples.(4)Referring to the existing methods of human image generation,this paper presents a method to generate human image by combining human appearance and three-dimensional human posture.In this method,a pseudo three-dimensional human model is constructed by combining the human appearance and three-dimensional human posture,and then the human image is generated by perspective projection and plane geometric deformation.In the INRIA dataset,the accuracy of the SVM classifier trained by the algorithm proposed in this paper with only 12 real images reaches 89.7425%,which is 7.8153% higher than that of 241 real images.In the OTB100 dataset,the SVM classifier trained by the algorithm proposed in this paper with only one real image has a classification accuracy of 91.0714% for similar image samples.So the human images generated by the algorithm proposed in this paper can effectively improve the performance of the image classifier when the training image data is insufficient.