From Classical to Deep Learning Methods in Image Downscaling

A Tutorial at the Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP) 2025

Overview

Image downscaling is a fundamental process in image processing that aims to reduce the spatial resolution of images while maintaining perceptual quality. It plays a vital role in applications such as image compression, visualization on low-resolution displays, and transmission over limited-bandwidth networks. The key challenge is to preserve important structural and textural information—such as edges and contours—while eliminating redundant data. This tutorial will cover a wide range of existing methods, evolving from traditional interpolation-based approaches to modern deep learning models, providing a comprehensive survey of the field.

Topical Plan (Tentative Duration: 3 Hours)

Time Topic Description
15 min Introduction & Motivation Why is image downscaling important? Applications, challenges, and the evolution of techniques.
60 min Part 1: Classical Methods
  • Interpolation-Based Approaches: Nearest-neighbor, Bilinear, Bicubic, and Lanczos resampling. Pros, cons, and visual comparisons.
  • Optimization-Based Techniques: Formulating downscaling as an inverse problem. Perceptual loss functions (e.g., SSIM) and gradient-based methods (L0 gradient minimization).
15 min Coffee Break -
60 min Part 2: Deep Learning Models
  • Adapting super-resolution networks (e.g., SRCNN) for downscaling.
  • Content-adaptive resampling kernels and learned resamplers.
  • Hybrid approaches combining classical interpolation with deep learning refinement.
20 min Part 3: Recent Advances & Future Directions Discussion on state-of-the-art structure-informed algorithms, performance on modern datasets (DIV2K, U100), and open research problems.
10 min Q&A and Conclusion Open discussion and wrap-up.

Relevant Publications & Materials

  1. G. Keys, “Cubic convolution interpolation for digital image processing,” IEEE Trans. Acoust. Speech Signal Process., vol. 29, no. 6, pp. 1153–1160, 1981.
  2. C. Lanczos, “Evaluation of a two-dimensional interpolation formula," Journal of the Society for Industrial and Applied Mathematics, vol. 12, no. 1, pp. 1–28, 1964.
  3. Johannes Kopf, Ariel Shamir, and Pieter Peers, “Content-adaptive image downscaling," ACM Transactions on Graphics, vol. 32, no. 6, pp. 173:1– 173:8, 2013.
  4. Nicolas Weber, Michael Waechter, Sandra C Amend, Stefan Guthe, and Michael Goesele, "Rapid, detail-preserving image downscaling.,” ACM Trans. Graph., vol. 35, no. 6, pp. 205–1, 2016.
  5. Sanjay Ghosh and Arpan Garai, “Image downscaling via co-occurrence learning,” Journal of Visual Communication and Image Representation, vol. 91, pp. 103766, 2023.
  6. Junjie Liu, Shengfeng He, and Rynson W. H. Lau, “L 0-regularized image downscaling," IEEE Transactions on Image Processing, vol. 27, no. 3, pp. 1076–1085, 2018.
  7. Wanjie Sun and Zhenzhong Chen, “Learned image downscaling for upscaling using content adaptive resampler,” IEEE Transactions on Image Processing, vol. 29, pp. 4027–4040, 2020.
  8. S. Ghosh, P. Nair, and K. N. Chaudhury, "Optimized Fourier bilateral filtering," IEEE Signal Processing Letters, vol. 25, no. 10, pp. 1555-1559, October 2018.
  9. S. Ghosh, R. G. Gavaskar, D. Panda, and K. N. Chaudhury, "Fast scale-adaptive bilateral texture smoothing," IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, no. 7, pp. 2015 - 2026, July 2020.

About the Presenter

Profile picture of Dr. Sanjay Ghosh

Dr. Sanjay Ghosh

Assistant Professor, Department of Electrical Engineering, IIT Kharagpur.

Email: sanjay.ghosh@ee.iitkgp.ac.in

Faculty Webpage | Google Scholar

Biography: Sanjay Ghosh received the PhD degree in Electrical Engineering from the Indian Institute of Science in 2019. Currently, he is an Assistant Professor in the Department of Electrical Engineering at the Indian Institute of Technology Kharagpur, India. His broad research interests are in computational imaging, brain signal processing, and machine learning methods for neurological disorder analysis. He received Best Student Paper Award at IEEE Global Conference on Signal and Information Processing (GlobalSIP) 2018 and Silver Award at International Conference on Biomagnetism (BIOMAG) 2022. Dr. Ghosh is a former fellow of “DAAD Postdoc-NeT-AI 2023” program, an initiative to collaborative research with German research institutions.