CosAE: A Novel Autoencoder for Super-Resolution Image Restoration using Fourier Series
2025-04-26
Researchers introduce CosAE, a novel autoencoder seamlessly integrating classic Fourier series with a feed-forward neural network. CosAE represents input images as 2D cosine time series, each defined by learnable frequency and Fourier coefficients. Unlike conventional autoencoders that lose detail in low-resolution bottlenecks, CosAE encodes frequency coefficients (amplitudes and phases) enabling extreme spatial compression (e.g., 64x downsampled feature maps) without detail loss upon decoding. Experiments on super-resolution and blind image restoration demonstrate state-of-the-art performance, highlighting CosAE's ability to learn a generalizable representation for image restoration.
Read more