Research On Ordinal Assessment Of Image Aesthetic Based On Deep Learning

The purpose of deep learning-based image aesthetics assessment is to enable computers to simulate the human cognitive process for image aesthetics,to narrow the gap between the aesthetic intelligence of computers and humans,and thus to build an image aesthetics assessment model.Using computable methods to predict human aesthetic perception of images has become one of the hot research problems in computer vision.This thesis investigates how to design an image aesthetic assessment model to extract to detailed features and holistic features related to image aesthetic assessment based on convolutional neural networks to improve the accuracy of image aesthetic quality assessment;it explores how to further refine the image aesthetic assessment classification problem by embedding the orderliness relationship between aesthetic labels into the classification problem from the image aesthetic binary classification problem to the image aesthetic ordered multiclassification problem.The main research is summarized as follows.(1)In this thesis,we propose an image aesthetic assessment method based on attention mechanism and holistic nested edge detection.How to effectively preserve the overall and detail characteristics of an image in the image aesthetic quality assessment problem is one of the key issues in image aesthetic assessment.Currently,most of the existing methods preserve the detail information of an image by random clipping,which leads to incomplete content and wholeness of the image.To overcome this drawback,we propose to extract both detail features and holistic features of an image using an attention mechanism and holistic nested edge detection: the attention mechanism is used to obtain the detailed features of an image associated with the aesthetic assessment of an image,and the holistic nested edge detection is used to extract the holistic edge features of an image.Our proposed method based on attention mechanism and overall nested edge detection for image aesthetic evaluation is performed on two datasets,and the accuracy of the proposed method for the image aesthetic binary classification problem is 88.13% and 94.10%,respectively;compared with other methods,the accuracy is improved with a maximum improvement of 10.74%.(2)In this thesis,we propose an image aesthetic grade assessment method based on ordered classification.Since the image aesthetic assessment problem is complex,it is difficult to meet the essential needs of the aesthetic assessment problem by merely dichotomizing images into high and low aesthetic quality,and there is an orderly relationship between image aesthetic classes.To solve the above problems,we propose an image aesthetic grade evaluation network based on ordered classification: further refine the image aesthetic classification and classify the aesthetic grades into four grades: low aesthetic,average,high,and high aesthetic;by adding a cumulative linkage model to the output layer of the image aesthetic assessment model,the ordered constraints are embedded in the convolutional neural network,which makes the difference between the prediction results and the real results smaller.Experiments were conducted on AVA and Photo.net datasets,and the MAE values obtained using the ordered classification-based image aesthetic grade evaluation method were 0.2360 and 0.2927,respectively.In comparison with other methods,the experimental results show the effectiveness of the proposed ordered classification-based image aesthetic grade assessment method in this thesis for the image aesthetic grade orderliness assessment task.