New inpainting compression uses derivatives for 2.76 dB PSNR gain
Researchers optimize not just pixels but derivatives and integrals for better image compression.
Inpainting-based compression works by storing a carefully selected subset of image data and reconstructing the rest via an inpainting process. Until now, the stored data has been almost exclusively pixel locations with grayscale or color values. This new work by Vassillen Chizhov, Ferdinand Jost, and Joachim Weickert (Saarland University, arXiv:2606.16581) introduces a general theoretical framework that allows incorporating arbitrary features describable by linear or nonlinear equations—such as derivatives of any order or local integrals. These features can be combined with linear or nonlinear inpainting operators (harmonic, edge-enhancing diffusion).
The key innovation is an algorithm that automatically optimizes both the location and the type of feature to store, turning the compression paradigm into a more flexible and powerful tool. Experiments show consistent quality improvements: with the same amount of stored data, using up to 5 different feature types yields average PSNR gains of 2.76 dB for harmonic (homogeneous diffusion) inpainting and 1.82 dB for edge-enhancing diffusion inpainting. This work opens the door to more efficient lossy image codecs that can store richer information without increasing file size.
- Proposes a general framework allowing arbitrary linear/nonlinear features (derivatives, integrals) for inpainting-based compression.
- Algorithm automatically optimizes both the location and type of feature to store.
- Using up to 5 feature types yields PSNR gains of 2.76 dB (harmonic) and 1.82 dB (edge-enhancing diffusion) with same storage.
Why It Matters
Improves image compression quality without increasing file size, enabling higher fidelity images in bandwidth-limited applications.