Font Size: a A A

Research And Implementation Of Video Style Transfer Stytem Base On Mobile Terminal

Posted on:2021-06-06Degree:MasterType:Thesis
Country:ChinaCandidate:B Y JiaFull Text:PDF
GTID:2518306308472364Subject:Software engineering
Abstract/Summary:PDF Full Text Request
Video style transfer algorithm is one of the important directions of neural style transfer algorithm.Because this kind of algorithm has high requirements for the computing power of computer equipment,it is difficult to deploy on mobile terminals.This paper studies the existing neural style transfer algorithm and proposes a real-time video stylization solution for mobile terminals.The solution improves the existing methods in terms of speed,flexibility and quality.At the same time,it designs and implements a video style transfer system for mobile terminals.The main research results of this article are as follows:First of all,in view of the problems of the existing stylized algorithm that requires a large amount of network computing and insufficient computing power of mobile terminals,the paper improves from the model level and proposes an image style transfer algorithm based on super-resolution architecture.Started with three aspects of the model structure to solve the bottleneck problem of computing power requirements.Secondly,in response to the problem of poor flexibility of the fast style transfer algorithm,the paper implements arbitrary style transfer,stylized process control and original image color retention functions through meta-network,double pyramid branch and color transfer technology,so that users can customize their own personal needs Adjust the stylized results.Finally,for the quality control of video style,the paper improves the existing video stylization scheme based on timing consistency constraints and introduces timing consistency constraints based on multi-level feature maps,which effectively improves the quality of video stylization results.The thesis carried out a lot of experimental work on the proposed algorithm and model,and compared it with other algorithms based on open source data sets.The experimental results show that the proposed algorithm and model have better performance.Based on the above model and algorithm improvements,the paper implements a real-time video style transfer system for server-side generation models and iOS-side deployment and operation.The style conversion process in the system is performed on mobile terminals,reducing the pressure on the server and network transmission.The paper introduces the system’s requirements analysis,summary design,and function module design.It conducts the system’s video stream acquisition and display module,stylized processing module,stylized process control module,style selection and information display module,and mobile terminal model file generation module.After detailed design and implementation,the functional test and performance test of the developed system were carried out.Test results show that the system can meet the operation requirements of real-time video style transfer on mobile terminals,and the actual operation results are good.
Keywords/Search Tags:Deep-Learning, Neural-Style-Transfer, Super-Resolution, Meta-Network
PDF Full Text Request
Related items