Research On Attribute Controllable Image Generation With Generative Adversarial Networks

The unsupervised controllable image generation task aims to fully explore the potential attribute pattern differences of the data and use them to guide the image generation process.Its performance is based on a full understanding of the structural features of the data,and is part of deep learning interpretable research,which is important for improving the stability and security of image generation models and promoting the application of artificial intelligence technology in more fields.The existing unsupervised controllable image generation methods based on generative adversarial networks generally have problems such as poor attribute control ability and low image generation quality.Thesis analyses and investigates the existing methods in terms of data potential attribute information utilisation and the rationality of attribute control constraints,and proposes a new method with better attribute control and better image generation quality.At the same time,thesis also presents a preliminary exploration of the application of unsupervised attribute controllable image generation methods in the field of continual discriminative representation learning.The main contributions of thesis are following.1.To address the problems that existing unsupervised attribute controllable image generation methods do not make sufficient use of real data and do not pay enough attention to the a priori information of attributes implicit in image data,thesis proposes an a priori knowledge-guided strategy to introduce attribute-related information latent in the data for the model through a self-supervised representation learning method,which successfully improves the accuracy of attribute control unlike existing methods.2.To address the problem that the constraint loss design of existing unsupervised attribute controllable image generation methods is unreasonable,which limits the attribute control effect and image generation quality,a new method of self-supervised attribute controllable image generation based on contrastive constraints is proposed.Through the feature similarity contrastive constraint,the differentiated attribute information of the generated image and the implicit coding of the controlled image generation are bound to each other,and finally achieve better attribute control performance and image generation quality.At the same time,the performance of the proposed method is further improved by combining realistic scenarios,proposing a contrastive less-labelled weakly supervised information introduction strategy and a model initialisation strategy,using no more than 1% labelled data and better model initialisation.3.Application research of unsupervised attribute controllable image generation methods in the field of continual discriminative representation learning is carried out.A generative replay training strategy is introduced to address the characteristics of sequential continuous discriminative representation learning tasks,and a joint incremental contrastive learning strategy and contrastive distillation strategy are further proposed to help the model maintain the discriminative ability of previous task data well in continual learning tasks,while achieving better inter-task attribute discrimination performance.The results show that the proposed method achieves better discriminative power than some supervised methods.