Semantically Encoding Enhancements for Unsupervised Color Feature Separation

With the advancement of deep learning and the increasing demand for high-quality data in AI research, the supervised learning and end-to-end paradigm has achieved tremendous success. Nevertheless, the black-box nature and limited generalization of end-to-end approaches impede their interpretability. Inspired by multi-modal learning, we propose a novel plural representation reductionist encoder-decoder framework that decomposes the original data into structurally differentiated latent variable combinations and uses a unified decoder to recover the original data. This framework aims to improve the interpretability and generalization ability of the model. Based on this framework, we design the Continuous-Discrete VQVAE (CD-VQVAE) model, which combines continuous latent vector representation with quantized latent vectors to achieve unsupervised separation of color features in images. The proposed model incorporates constraints on unitized codebook embedding vectors and introduces orthogonal constraints between pairwise feature layers of discrete latent vectors, resulting in a structured latent space that reduces data redundancy. Experimental results demonstrate that CD-VQVAE achieves unsupervised color feature separation and effectively improves the generation quality of the model.