A Study on Performance Improvement of DenseNet-Based VAE model Using Self-Attention and Additive Fusion

This study proposes an improved method to simultaneously address the three key bottlenecks of Variational AutoEncoders(VAEs): low-level texture loss, lack of global context, and latent space information degradation. The proposed model combines DenseNet-based encoder architecture with an additive fusion technique leveraging self-attention to enhance image reconstruction performance. The encoder employs DenseNet-121 to facilitate feature reuse across layers, effectively preserving local details, while the self-attention mechanism in the decoder learns long-range dependencies. Features extracted from both paths are dynamically integrated via additive fusion, enabling balanced reconstruction of global structures and high-frequency textures. Experiments conducted on the CIFAR-10 dataset demonstrate that the proposed model outperforms conventional VAE, achieving improved PSNR, SSIM, MSE, and LPIPS metrics of 20.93, 0.5886, 0.008733, and 0.1722, respectively. Additional experiments on the ImageNet-R dataset show that the proposed model consistently surpasses standard VAE performance across all metrics. Although it exhibits slightly inferior high-frequency texture reconstruction compared to Vector Quantized Variational AutoEncoder(VQ-VAE), the model demonstrates competitive performance as a continuous latent space-based approach. This study also validates reconstruction quality across varying resolutions and styles, and suggests future directions including scaling to high-resolution datasets, model compression, and hybrid quantized latent representations.