Research On Automatic Generation And Aesthetic Quality Evaluation Of Advertisement Layout Images

With the opening of online and offline services,the retail industry has ushered in a new round of rapid development.This is not only the rapid growth of the design demand for advertising,especially for web banner advertising,but also a higher demand for the timeliness of advertising design.At present,advertising design is mostly completed by professional designers.Due to the high dependence of the whole design process on human,especially the selection and processing of image materials,it is difficult to improve the design efficiency quickly to meet the explosive growth of business needs.However,the development of material data and artificial intelligence technology on the Internet provides an opportunity to solve this bottleneck.This thesis mainly studies the automatic generation of the layout of advertising images and the aesthetic quality evaluation of the generated layout images.The main work of this thesis is as follows:First of all,in order to realize the automatic generation of the layout of advertising images,this thesis proposes a network of layout generation based on Generative Adversarial Networks.In this work,we do not consider the specific content of ad descriptors,only the location of ad descriptors in the whole image.Firstly,according to the particularity of advertising images,the model will consider the visual information and attribute information of advertising images,including image style,product classification and suitable people.Then,the two kinds of information are encoded separately,and then integrated as the input to Generative Adversarial Networks.At the same time,according to the two kinds of information,the positions of the main image and the description language of the advertisement on the whole image are sampled in the data set.Then,the network generates the layout images according to the sampling information,and calculates the similarity score using L2 distance,and then outputs the top three images.Finally,according to the experimental results and data analysis,the model in this thesis can well complete the generation of advertising layout images and the quality of the generated images is good.Secondly,in order to improve the quality of the generated layout images,this thesis proposes an aesthetic quality evaluation network of advertising layout images,which includes Multi-Attribute Feature Network,Attention Network and Language Generation Network.First of all,5 professional advertising designers and 10 advertising industry practitioners are invited to grade the layout images.The scoring options include composition,color,image focus and overall impression,so as to construct a special layout scoring data set.The data set contains the numerical scores of four aesthetic attributes and the linguistic evaluation of each aesthetic attribute.Then the Multi-Attribute Feature Network calculates the feature matrix of different attributes through multi-task regression of four attribute scores,and the Attention Network dynamically adjusts the attention weight of channel dimension and spatial dimension of the acquired features.Then,the Language Generation Network generates image subtitles through the long-term memory network.The long-term memory network needs the real content of language evaluation about each aesthetic attribute in the data set and the feature mapping after the network adjustment.Finally,through the experimental analysis,the model of this thesis can get a good score of each image's aesthetic attributes and language evaluation,at the same time,it can also get the aesthetic score of the whole image.