Image Diversity Generation Based On Generative Adversarial Network

The purpose of image diversity generation is to transform a given image into a variety of target images that are different from the original image on the premise of maintaining the o-riginal main features through certain processing.This technology is widely used in creative design,partial generation and image editing,etc.It plays an extremely important role in the artificial intelligence,mobile Internet and multimedia field.The traditional shallow im-age diversity generation algorithm mostly adopts linear mapping,which simply splice and transform the image to generate new images,so it is difficult to dig out the hidden depth nonlinear features in the image,and damages the original spatial structure of the image.Re-cently,image diversity generation algorithms based on deep learning gradually extract fea-ture information that can describe the essential content of images through the hierarchical learning,which has become a hot spot in current research.Especially the rise of generative adversarial network,because of its excellent autonomous generation ability and ingenious weakly supervised learning mode,has made a qualitative leap in image data generation speed and quality.However,the images generated by existing image diversity generation based on generative adversarial network algorithms are too similar,and multiple different inputs correspond to similar outputs,that is,mode collapse.Target at these drawbacks,this paper starts with deep learning and adversarial learning,and further researches the image-based random diversity generation and the designated diversity generation algorithms.The specific research contents are as follows:(1)For the exsiting image random diversity generation algorithm,they mainly consider the error relationship between the generated samples and the original samples with noise and redundant features,which lead to the problem of blurred images and insufficient diversity.This paper proposes a Consistent Embedded Generative Adversarial Network(CEGAN).The network uses subspace learning to extract the distribution of style features,and random-ly generates the learned distribution to achieve generation diversity.At the same time,it uses latent spatial adversarial learning to suppress the adverse effects of noise and redun-dant features in the original image to achieve high quality generation.The qualitative and quantitative experimental results on five commonly used databases show that CEGAN has excellent performance.(2)Target at the problem that the existing image-specific diversity generation algorithm-s are susceptible to interference from the original attributes in the image when generat-ing the specified attributes,an Image Diversity Generation Algorithm based on Joint At-tribute Learning(JAL)is proposed.The algorithm combines the attribute decomposition network and the attribute classifier to decompose the image into attribute-related features and inherent-structural features.At the same time,it further updates the network to obtain a better feature description via attribute loop consistency reconstruction,so that the proposed method can be guided to generate high quality images with specific attributes when giv-en the attribute.The experimental results on RaFD and C elebA datasets demonstrate the effectiveness of the proposed JAL.