Computability Design Of Some Semantic Features And Quality Assessment For Image Aesthetics

With the rapid development and popularity of mobile electronic devices such as smart phones,the era of universal photography has entered.Due to the differences of photography conditions and the aesthetics appreciation ability of the photographer,the aesthetic quality of images also shows a big difference.In order to meet people's demand for image appreciation,it is more and more important to evaluate the quality of the image from the perspective of aesthetics.The technology can be used in areas such as image retrieval,automatic image editing,personal photo album management,and human-computer interaction.Due to the subjectivity and complexity of human aesthetics,it is a challenging task to quantify the evaluation criteria of aesthetic quality of images.This thesis is devoted to designing computable features with aesthetic semantics and proposes new feature extraction algorithms.On the basis of this,the handcrafted features and the deep features are complementarily combined,and an algorithm for scoring the image aesthetic quality is proposed.The main work of this paper is as follows:(1)In image aesthetic quality assessment,for the first time,we consider the influence of line order on the evaluation of image aesthetic quality,and design and give the characterization of line order feature and calculation methods.A method based on ordered line characterization is proposed,and the angle of statistical lines is transformed into the gradient angle of pixels in the statistical image.The gradient angle histogram is constructed as the line order feature.(2)Propose a new clarity contrast feature for the salient region of images and give the corresponding calculation method.The regional clarity index is constructed by measuring the clarity of each pixel in the image.Firstly,the gradient information of each pixel is calculated,and the pixels in the salient region and the background region are summed respectively by the HVS filters to obtain the definition indexes of the two regions.Then the clarity contrast feature of the image is obtained.This feature is not only easy to calculate,but also can better suppress white noise,reflecting the true perceived clarity of the image.(3)Analyze the complementarity between the handcrafted features and the deep features of image aesthetics,and propose an image aesthetic quality scoring algorithm with complementary features.Handcrafted features make it difficult to describe the aesthetic information of the images in a comprehensive way.The deep features learned by the convolutional neural network make it difficult to analyze the specific regions of the images and their correlations.Both of them are complementary in the description of the image aesthetics information.In this paper,we trains PARN network to extract deep features,and the SVR regression algorithm is used to establish the image aesthetic quality scoring model for each of the two types of features,and the obtained two groups of scores are weighted and summed to obtain the final aesthetic score.In addition,aiming at the above algorithm,a weight calculation method based on spearman rank correlation coefficient is proposed.(4)The above algorithm process is implemented,and comparative experiments are carried out on two data sets of AADB and AVA.Experimental results show that the aesthetics semantic features of artificial design in this paper not only enhance the performance of the model but also complement the other artifacts when combined with the deep features.The ranking of the images based on the aesthetic quality scores is in good conformity to subjective assessment results.