Research On Linear Reconstruction Based Representation Learning And Its Applications In Images Analysis

With the advent of the information age,the computer has become an essential information process-ing tools in the modern human society.As modern communications technology advances and popularity of the Internet,the image is becoming most common information carrier in people's daily life.Com-pared to the traditional information carrier,that is the word,the image has obvious more advantages:intuition—images are direct records of the scene.Comprehensiveness—images could reproduce the scenes exhaustively.Versatility—images are not restricted by national borders and language.Con-venience—images are easier to understand.Therefore,automatic image analysis and processing via computer is the basis of the intelligent community.Different from the word processing,automatic image analysis and processing is of a greater chal-lenge.First,people lack the use of understanding on the image processing and analysis by computer.Using and handling of the text has thousands of years of history and human have accumulated a wealth of experience.But the emergence and development of digital images and computer science are no more than a hundred years.How to use the computer science and technology to extract information according to the inherent characteristics of the image has always been the target of computer vision.Secondly,people lack the guidelines on using computer for image processing and analysis.Currently,the operation of human image recognition system are referenced for designing automatic image analysis and processing systems.However,the human image recognition system has experienced a few million years of evolution.It is a very complex system.Compared to it,the history of the development of com-puter science is too insignificant.How to effectively grasp the essence of the human image recognition systems and use it for image processing will be a long way for the development of computer vision.According to the existing image representation methods,the features of images are extracted mainly based on two aspects:imitate the human image recognition organ to extract the internal structure of the image—feature descriptor;imitate the human nervous system to process the images—shallow and deep learning.In this paper,based on a basic application of image analysis,that is the face recog-nition system,and starting from shallow learning,methods on face alignment and image representation in facial recognition system are improved via the linear reconstruction.Specifically,the contributions of this paper are as follows:(1)A new manifold learning method is proposed and it is applied in face alignment.Currently,shallow learning usually assumes the spatial structure of the samples is linear.Although this could reduce the complexity of data processing,the topological structure of the data has been overlooked.In fact,the high-dimensional data usually has a manifold structure,the most obvious example is the facial shape vector space.However,in the parameter models of face alignment,the shape model assumes the facial shape vector space is linear.Manifold learning as a non-linear embedding method can effectively embed high-dimensional data into its manifold space non-linearly,resulting in a linear structure of the data.But due to the estimation of the dimension of manifold space,the calculation complexity is too great to fulfill real-time.By smoothing the local submanifolds,a new manifold learning methods is proposed based on the local tangent space alignment.Because of its explicit projection and the manifold transform in the original space,it could be well combined with the shape model in face alignment.So the manifold of facial shapes can be embedded into the model.(2)An improved spatial non-negative matrix factorization is proposed.In the representation learning methods based on linear reconstruction,non-negative matrix factorization is a specialized fea-ture learning method designed for image data.Compared with the previous representation methods,the basis images of non-negative matrix factorization have better local structures.So non-negative ma-trix factorization as a parts-based representation learning method,the representation vectors are more robust and understandable.In order to further improve the basis image of locality,the improvements on non-negative matrix factorization are mainly focused on embedding the spatial information into the model.However,such information is usually derived from the two-dimensional spatial network of the image,so they are the lack of the dependency on the content of data.Hence,according to the factor analysis on the relationship between image features,a new spatial regularization which merges the data distribution and spatial structure is proposed.Then the max-margin constraint is combined with it.It not only achieves the embedding of space structure,fusion of the discrimination and locality,but also reduces the contradiction between discriminant constraints and local constraints on the representation vectors.(3)A new attribute feature is proposed.Compared with the traditional feature descriptors,attribute is a higher level feature.It is not about some geometrical structure in the images,but a semantic information reflected in the image.Because of this,attribute has better explanatory for human.However,definition of the semantic is various,and many semantics are relatively abstract.So attribute learning is usually complicated and inaccurate.Specifically,for the continuous attribute,appropriate feature were extracted for each attribute and each attribute has their own attribute classifier.The outputs of the classifiers are used for attributes.Which attributes are utilized for data representation,which feature fits the characteristics of each attribute and the differences of the design on attribute classifiers will all affect the quality of attributes.Thus,based on prototype theory in Psychology,a class relative attribute is proposed—prototype based relative attribute.Where,each attribute separately reflects the correlation with the classes of given samples instead of choosing the problems-depend attributes.Meanwhile,each attribute uses the same features for attribute learning.So the attribute learning is simplified to a certain extent.