Adaptive image attribute editing model and editing method based on classification adversarial network

Abstract

The invention provides an adaptive image attribute editing model based on a classification adversarial network, and the method achieves the accurate attribute conversion and high-quality image generation functions through the construction of an upcoiler residual network and the addition of an attribute adversarial classifier Atta-cls in a discriminator. A decoder is constructed by adopting an upper convolution residual error network Trresnet, attribute features and content features are selectively extracted, the problem of limitation of jump connection in a deep encoder decoder structure is solved, the attribute features of a target image are enhanced, a more accurate and high-quality image is generated, and the performance of a model is improved. Under the influence of the idea of the generative adversarial network, an attribute adversarial classifier Atta-cls understands the deficiency of the converted image in an adversarial learning mode for the attribute difference, and further optimizes the converted image according to the deficiency. According to the invention, the assessed attribute label is enabled to approach the source label through the attribute continuity loss function, and the attribute continuity of the generated image is ensured.

Description

technical field [0001] The invention belongs to the technical field of attribute editing of image generation, and in particular relates to an adaptive image attribute editing model and editing method based on classification confrontation network. Background technique [0002] Attribute editing, also known as attribute transformation, aims to change the attributes of an image, including one or more attributes of hair color, gender, style, etc., while keeping other attributes unchanged. The key to attribute editing is to achieve accurate attribute conversion and generate high-quality images. In recent years, Generative Adversarial Networks (GANs) have greatly promoted the development of attribute editing. Generative Adversarial Networks (GANs) are defined as a minimax game with a generator and a discriminator in which the generator generates images that are as realistic as possible and the discriminator tries to distinguish the synthesized image from the original image. GAN ...

AI Technical Summary

Problems solved by technology

However, the existing image attribute editing methods only use the original image as the input of the attribute classifier when training the classifier, and use the optimized classifier to improve the generator, ignoring the influence of the generated image on the accuracy of enhanced attribute transfer. Difficult to spot attribute differences between generated and real images

Images