关键字：-

摘要：There exists a general problem in numerous disentangled representation learning algorithms that improves disentanglement performance by sacrificing generation performance. In order to solve the contradiction between disentanglement and generative performance, we propose a high-fidelity synthetic causal disentangled representation learning framework (HCAPE) combining Variational Auto-Encoder (VAE) and Generative Adversarial Network (GAN). First, in the inference stage, the assumption that the latent factors of the iation are mutually independent is abandoned. And a Structural Causal Model (SCM) in the encoder is resorted to learn causal representations. Second, a supervision term and balance parameter eta were introduced to reduce the interference of disentanglement performance on generation performance. Finally, in the generation stage, a discriminator estimates the model's gradient information, restoring high-frequency details in the image, thus synthesizing high-fidelity images. Experimental results show that the causal representations learned by the model are semantically interpretable and can generate counterfactual data. Most importantly, the model is able to generate high-fidelity images, achieving an optimal balance between disentangling and generative performance.

卷号：265

期号：-

是否译文：否