Application And Research Of Panda Image Automatic Generation Technology

Image generation is an important branch of the current computer vision field,and it is also a hot research topic in recent years.On the other hand,panda culture is a symbol of Chinese culture and plays a pivotal role in international culture.In the emerging digital entertainment industry,mass creators have higher thresholds for the design of panda images,and it is difficult to restore their cuteness and cuteness.Therefore,the proportion of panda culture in the industry is relatively low.In view of the difficulties in the creative process,this article proposes a method for generating panda images by line drawing based on the existing image generation technology based on the requirements and creative process of entertainment and creation,and provides intuitive assistance for creators.This article's main work has the following three points:In response to the lack of richness of panda image data in the existing market environment,a segmented image data set of pandas was produced through deep learning methods,which visually displayed the individual image of pandas,eliminated the interference of background and environmental factors,and provided subsequent image generation and other work service.This paper also improves a non-end-to-end phased image generation algorithm,which combines edge generation and image generation algorithms to generate a panda image based on input sparse edges.Because the two algorithm models are completely separated during the training process,they can be repeatedly trained out of order on the basis of a small amount of pre-training until the effect is perfect.At the same time,this paper proposes an improvement method for the loss function,training method,and model structure in the generated model,and proposes an image optimization function for segmented image data sets to improve the reality of panda image generation.This method has the following advantages: 1.By first repairing the edges and then generating the image,the panda image is gradually generated,which is in line with the characteristics of entertainment creators when they first draw lines and then fill colors.2.Converted sparse edge generation and image generation problems to image patching problems 3.A large number of cutting-edge generation adversarial networks such as residual networks and spectral normalization were used in the model,so that model training can avoid pattern collapse during the training process.This paper proposes an image generation system based on the giant panda resource platform,which provides users with an algorithmic experience through the aforementioned panda image generation method.On the other hand,the system integrates image segmentation models to provide users with online segmentation functions.According to the feature extraction network module,the user can retrieve a large number of pictures in the database by using a map search method.