Research On Automatic Layout And Evaluation Of Presentations Based On Deep Learning

Presentations may be the most popular way for efficient communication in our daily life these days,whose layouts significantly affect the performance of information exchange.There are many disadvantages to manual layout design,which is time-consuming,unaesthetic,unprofessional,etc.To meet the increasing demand of content-aware information presentation,automatic layout design has got great attention and is still a challenging task.This paper addressed both the layout quality evaluation and automatic layout design of presentations.A subjective evaluation algorithm of layout quality based on convolution neural network,an objective evaluation criteria of layout quality,and an automatic layout design algorithm based on deep convolution generative adversarial network were presented.The genetic algorithm was then integrated with the automatic layout design process,which further optimizes the layout results.The main work is as follows:(1)The convolution neural network is used to predict the subjective evaluation for layout quality based on human evaluation results.Each layout sample is mapped into two-dimensional grids,whose categories and attributes are fed into the neural network as the input,and the corresponding evaluation result is taken as the output.Further,the accuracy of evaluation results under different tolerances is discussed.(2)The objective evaluation criteria of presentation layout quality are defined based on some aesthetic principles and layout constrains.The energy-based model,mainly based on the aspects of alignment,balance,white space,scale,overlap and boundary,is used to measure the overall quality of the layout.In the model,different criteria their own energy items with different weights.(3)One automatic layout design algorithm was introduced,which is based on the idea of Layout GAN and also takes the attributes and topology constraints of page objects into account.The attribute extraction algorithms and network training process were also introduced.Due to the too long training time of the Layout GAN algorithm,an automatic layout algorithm based on DCGAN was designed and implemented,which could speed up the training process obviously.(4)In order to reduce the randomness of the above generative adversarial network based algorithms,with the random initialization,and making use of the genetic algorithm,the subjective and objective evaluation results were used as feedback to optimize the layout results gradually.Final layout results outperform the ones not optimized.Experiment results shew that the subjective evaluation network performed well and the evaluation results were consistent with human's.The layout quality generated by the improved Layout GAN was obviously improved both subjectively and objectively.The training efficiency of DCGAN was significantly improved and has comparable layout results as the improved Layout GAN.The layout results of the genetic version could obtain obvious local optimization effect in subjective and objective evaluation.