Research And Application Of Image Aesthetic Quality Evaluation Based On Transformer

The aesthetic quality evaluation of images is a study that uses computers to simulate human visual perception and then quantitatively evaluate the aesthetic quality of images.As the task requirements for image aesthetic analysis gradually increase,traditional aesthetic assessment models have certain limitations.At the same time,with the vigorous development of graphic design and data visualization technology,the aesthetic quality assessment of unnatural images has gradually become one of the mainstream tasks in the field of aesthetic assessment.In addition,aesthetic quality assessment for specific pixel images composed of single elements is also a challenging issue at present.This paper explores the aesthetic quality evaluation of images based on the Transformer architecture.The assessment scenarios include natural images,data visualization chart images,and pixel images.Two large-scale aesthetic datasets with multi-category aesthetic labels and three feasible end-to-end aesthetics are proposed.The model is evaluated and the effectiveness of the model is verified under three different scenarios.Firstly,this paper researches the aesthetic quality evaluation of natural images.Previous research work mainly focused on how to explore the mapping relationship between images and aesthetic annotations using convolutional neural networks.When the image information is complex and the background information is rich,it is necessary to correlate the aesthetic characteristics of different regions on the image.Due to the lack of advantages of convolutional neural networks in remote modeling,previous researchers have not fully explored the aesthetic weights of different regions on an image.Based on the Transformer architecture,this paper builds a model IAFormer for natural image aesthetic quality evaluation and proposes an interval attention module suitable for natural image aesthetic evaluation.In application,the model proposed in this paper can not only effectively evaluate the aesthetic quality of natural images,but also provide effective information references for aesthetic clipping tasks based on the aesthetic weights of different regions on the image.This paper also researched the aesthetic quality evaluation of data visualization charts and images.The previous visualization chart image evaluation work lacked a rich and complete visualization chart aesthetic dataset,and the aesthetic evaluation model was also incomplete.This paper constructs a data set CADataset for visualization chart aesthetic evaluation and proposes a novel end-to-end visualization chart image aesthetic quality evaluation model CAFormer based on the Transformer architecture.In application,the model proposed in this article has a strong ability to capture the regional details of visualization chart images,which can effectively evaluate the aesthetics of data visualization chart images,and provide effective reference information for optimizing the style of visualization charts.Finally,this paper conducted a study on the aesthetic quality evaluation of pixel images.In the past,pixel image research has mainly focused on the generation of pixel images.Most aesthetic quality evaluation strategies for pixel images follow objective image evaluation methods,ignoring the subjective aesthetic preferences of user groups.At the same time,due to the lack of pixel image aesthetic data sets with aesthetic annotations,there are certain bottlenecks in the aesthetic evaluation of pixel images.This paper constructs PADataset,a pixmap aesthetic dataset with rich annotations,and introduces PAFormer,an end-to-end pixmap aesthetic evaluation model based on Swin Transformer.In applications,PAFormer can not only effectively evaluate the aesthetic quality of pixel images,but also evaluate the aesthetic quality of pixelated natural images,providing effective auxiliary information for image stylization tasks.