Upsampling
The process of increasing the resolution of an image or signal by adding pixels and estimating missing information through interpolation or deep learning techniques.
Upsampling refers to increasing the pixel count of an image beyond its original resolution. When scaling a 256x256 image to 1024x1024, approximately 93.75% of pixel data must be synthesized through estimation.
- Nearest-neighbor: Copies the closest pixel value. Produces blocky artifacts but preserves hard edges, ideal for pixel art. Negligible computational cost
- Bilinear: Weighted average of four nearest pixels. Smooth results but blurs edges
- Bicubic: Cubic polynomial fitted to 16 surrounding pixels. Photoshop's default upscaling method, balancing sharpness and smoothness
- Lanczos: Windowed sinc filter maintaining sharpness with minimal ringing. Used in video processing via
ffmpeg -sws_flags lanczos
Deep learning super-resolution (SRCNN, ESRGAN, Real-ESRGAN) has become the state of the art, learning high-frequency patterns from training data to reconstruct details classical methods cannot recover. Modern models achieve PSNR above 30 dB at 4x magnification.
In practice: bilinear for real-time rendering, bicubic or Lanczos for print, and AI super-resolution when quality is paramount.