Performance Comparison of Autoencoders Using Multi-Head and Skipping Connections
Gyeongmin Kim, Suyeon Lee, Jinhwan Koh · 2024
This study introduces a novel Autoencoder design that enhances the conventional CNN-based Autoencoder architecture for more effective image noise reduction. By incorporating multi-head Autoencoders into the U-Net structure in a parallel configuration, this new architecture demonstrates approximately 1.25 times better Peak Signal-toNoise Ratio (PSNR) compared to traditional Autoencoders, proving its superior ability in reducing image noise.