基于深度学习的广告布局图片美学属性评价-《计算机技术与发展》

文章信息/Info

Title:: Aesthetic Attribute Evaluation of Advertising Layout Images Based on Deep Learning

Author(s):: SHEN Jia-min; BAO Bing-kun; School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications, Nanjing 210000,China

Keywords:: advertisement layout images; aesthetic evaluation; aesthetic attribute title and aesthetic score; deep learning; multi-task learning

摘要:: 图像质量美学评价是近十年来比较热门的课题,但是研究的大多是对自然图像的美学评价。然而随着互联网技术的发展,线上广告业务得到了迅速发展,因此准确高效地评价一张广告布局图片的好坏是很有必要的。所谓广告布局图片,即广告图片不考虑广告语的具体内容。为了研究广告布局图片的质量美学评价,引入了一个新的数据集 ALID,该数据集包含了四个美学属性的数值评分和语言评价;提出了美学多属性网络,该网络包含了三个部分:多属性特征网络、注意网络和语言生成网络。多属性特征网络通过 4 个不同的美学属性得分的多任务回归计算不同属性的特征矩阵,注意网络动态地调整所获特征的维度,最后语言生成网络通过长短期记忆网络生成图像字幕。实验结果表明,根据图像字幕的评价标准,该文设计的模型优于传统的 CNN-LSTM 模型和现代的 SCA-CNN 模型。

Abstract:: The aesthetic evaluation of image quality is a hot topic in the past decade, but most of the research is about the aesthetic evaluation of natural images. However,with the development of Internet technology, online adv-ertising business has been developed rapidly,so it is necessary to accurately and efficiently evaluate the qu-ality of an advertising layout images. The so-called advertisement layout images mean that the advertisement images do not take into account the specific content of the advertisement language. In order to study the aes-thetic quality evaluation of advertising layout images,a new data set,ALID,is introduced,which includes the nu-merical evaluation and language evaluation of four aesthetic attributes. An aesthetic multi-attribute network is proposed,which consists of three parts:multi-attribute feature network,attention network and language gene-ration network. The multi-attribute feature network calculates the feature matrix of different attributes through multi-task regression of four different aesthetic attribute scores. Attention network dynamically adjusts the dimensions of acquired features. Finally,the language generation network generates image subtitles through the long short-term memory. The experiment shows that the proposed model is superior to the traditional CNN-LSTM model and the modern SCA-CNN model according to the evaluation criteria of image subtitles.