| The conditional generation of high-resolution images can support the creation of image contents,which is used in scenarios such as advertising marketing,art design,and enhancement of scarce datasets in machine learning tasks.This study focuses on the high-resolution image generation technology with semantic labels as conditions.In this study,a high-resolution image generation model is proposed on the basis of the conditional generation adversarial networks.By injecting multi-level label condition information into the generation module,the model is guided to generate images that highly match the semantic labels;the multi-scale discrimination module guides the model to generate high-resolution images with high quality at multiple scales.Finally,an object feature coding network is introduced on the basis of the adversarial model,so that the generated images support object-level semantic editing.The main work is summarized as follows:(1)The multi-level label constrained generator improves the matching degree of the generated images and the semantic labels: the image contents generated by the condition and the object represented by the corresponding semantic label should be semantic consistent.The multi-level label constrained generator proposed in this study supplements the lost semantic condition information by connecting the semantic labels of corresponding scales to the decoder network layer,thereby improved the matching degree of the generated images and the semantic labels.(2)Multi-scale feature discriminator improves the definition of generated highresolution images: for the task of high-resolution image generation,this study uses a multi-scale feature discriminator composed of multiple basic discriminators of different scales,discriminating high-resolution images from multiple feature scales forces the generator to generate images with higher quality on multiple scales.In addition,this article uses a new loss function composed of Hinge Loss,Feature Matching Loss,and Perceptual Loss.This combination of loss functions can stabilize the training process of the model and further improved the clarity of the generated images.(3)The object feature coding network realizes the semantic editing of the generated images: the object feature coding network can generate a feature code of the corresponding style for each semantic object in the real image.Thanks to the improved semantic matching and clarity of the generated images,the model can use different semantic label and feature code inputs to generate clear images with different contents and styles. |
