Research And Implementation Of The Background De-Distortion Method For Beauty Image Based On Generative Adversarial Network

With the development of economy and society,the innovation of Internet technology,and the improvement of smart phone hardware capabilities,people have become accustomed to communicating with friends through images on the online social media platforms,and they are becoming more demanding of the aesthetics of self-images.As a result,a lot of image-beautify applications have been developed,which only need users to perform simple drag operations on mobile devices to achieve intelligent body beautification.However,these applications often make the portrait and background in the image partially distorted and unnatural,which in turn degrades the image quality significantly.This thesis proposes a stacked semantically-guided framework based on the generative adversarial network,which can capture and restore the distortions around the humans and the adjacent background effectively.The main contributions of this thesis include the following aspects:Firstly,we implement a crawler algorithm,which can automatically crawl large-scale real person images on the Internet,and crawl 116054 samples.Secondly,we construct a largescale image distortion dataset.To construct the dataset,we design and implement an automatic intelligent body beautification algorithm by combining the deep learning algorithms and the traditional machine learning algorithms.Thirdly,we train the proposed stacked semantically-guided learning network with the constructed image distortion dataset so that the network is able to discriminate and restore the distortion regions of the humans and background in the image.On the constructed large-scale image distortion dataset and realworld character image dataset,the proposed algorithm in this thesis achieves the best visual and quantitative results compared with other advanced deep learning algorithms.At the same time,we also verify the effectiveness and generalization of our algorithm through the sufficient ablation experiments and comparative experiments.