Visual Comparison Based On Relative Attribute Learning

Visual attributes are the basic feature of image,while the attributes and relative attributes based learning has become research focus in computer visual application.Because computer is expressed in binary data,the linguistic information transmitted by human beings is ever-changing so that there is a "semantic gap" between the underlying data information and the high-level semantic information.Visual attributes,which are mid-level properties that appear across category boundaries and often vary in terms of their perceived strengths.They are properties observable in images that have human-designated names(e.g.,'pointy','smiling'),so they can capture more semantic relationships to solve the "semantic gap".Therefore,a lot of research work on visual attributes came into being,but the main research is whether an attribute exists or not in recognizing object,that is,binary attribute.In recent years,the study based on relative attributes has emerged in endlessly.Aside from using global ranking function model to learn relative attributes,there are also local learning methods for fine-grained visual comparison.Subsequently,a large number of researchers have proposed a relative attribute model based on deep learning.Although these studies have achieved good experimental results,they result in higher time complexity and space complexity due to the structural characteristics of deep learning.Aiming at the limitation of current visual comparison model based on relative attributes,The contents of this thesis are as follows:for the general visual comparison,a more accurate global visual comparison model is proposed;for the fine-grained visual comparison,a more accurate local visual comparison model is proposed;for the ordered or similar visual comparison models,a similar visual comparison that can simultaneously predict more,less or equal visual comparison relations is proposed.The main work and innovations of this thesis include:For the global visual comparison model,there are some unrelated image pairs for training samples,which lead to the deviation of visual comparison methods for relative attributes.This thesis presents a method based on Linear Regression Model(LRM)and Linear Discriminant Analysis(LDA)for global visual comparison based on relative attributes.Recent research is almost about the relative attribute model based on Ranking SVM,which is sensitive to support vector.If there are some unrelated image pairs in the training samples,and they are selected as support vectors,the erroneous Ranking SVM model function will be generated eventually,which will lead to a great discount for visual comparison.Linear regression can solve the problem of support vector sensitivity well,and the visual comparison model based on relative attributes can be established by using linear regression function to obtain effective visual comparison.In order to prevent the over-fitting of the parameters of linear regression function,we use ridge regression to regularize the loss function of linear regression.In addition,in order to avoid the problem of over-fitting caused by high-dimensional features,low-dimensional and discriminant features can be obtained by using LDA method to reduce the dimension of features.Experimental results on the three databases demonstrate the advantages of the proposed method.Aiming at the problem that global features in fine-grained visual comparison can not express some attributes with local characteristics and metric learning methods run for much time,this thesis mainly proposes a new metric learning to find K-nearest neighbor training image pairs for each test image pair,and a feature representation that can capture global and local information simultaneously.We need to learn a discriminant metric learning method called Relative Attribute Quadratic Discriminant(RQDA)method which can simultaneously reduce dimensionality and conduct metric learning.In addition,the Histograms of Oriented Gradient(HOG)descriptor has been widely applied for robust visual object recognition,which describes fine-scale gradients,fine-orientation binning,relatively coarse spatial binning,and high-quality local contrast normalization in overlapping descriptor blocks.Therefore,we propose an informative feature descriptor of combining HOG with gist to take into account of both global and local features.Experimental results demonstrate the proposed method is an effective fine-grained visual comparison methodAiming at the problem that only ordered image pairs can be detected but not similar image pairs in ordered or similar visual comparison,a one-versus-one multi-class classification model is proposed to predict which image in an image pair has stronger,weaker or similar visual attribute strength simultaneously.Meanwhile,the LDA model is used to reduce the feature dimensionality to overcome the over-fitting caused by high-dimensional feature.Alth ough prior methods can predict which image in an image pair has stronger or weaker attribute strength,they cannot predict the case similar attribute strength.To solve this issue in local learning method,a method to realize Just Noticeable Differences by using parzen window probability density function was proposed,which could predict whether the attribute strength is similar or not.The proposed method can simultaneously predict which image in an image pair has stronger,weaker or similar visual attribute strength.Experiments show that the proposed method is an effective method for ordered or similar visual comparison.