GB-LSR: A Fast Local Spectral Image Representation with a Single Global Bandwidth for Continuous Reconstruction and Super-Resolution
Researchers have introduced GB-LSR, a novel local spectral image representation designed for efficient and continuous image reconstruction and super-resolution. This method partitions images into patches, each with coefficients derived from shared convolutional features and a single global bandwidth scalar. GB-LSR demonstrates superior performance in native reconstruction benchmarks compared to existing methods, achieving higher PSNR and LPIPS scores while operating at a significantly faster inference speed. The approach also shows promise in arbitrary-scale super-resolution tasks, offering competitive results with improved speed and reduced memory usage. AI
IMPACT This new representation could lead to more efficient and higher-quality image reconstruction and super-resolution tools.