Curated References & Further Reading

"I have ingested every image processing paper since 1962 and retained mostly the ones with good figures. My advice: one good book, one good course, and the humility to read the documentation before the fourth hour of debugging."

A Bibliographically Overextended Vision Model

The previous section (Section 8.3) equipped you to evaluate methods; this one equips you to keep learning them. Part I compressed sixty years of image processing into eight chapters, and every one of those chapters has a deeper literature behind it. The problem is no longer scarcity but routing: given a question, which shelf do you pull from? Figure 8.4.1 shows the routing logic this section follows.

1. The Bookshelf Beginner

Four books cover Part I's territory, each from a different angle, and Table 8.4.1 maps them onto the chapters you have read so you can jump straight to the right pages.

For video, three free courses stand out. Shree Nayar's First Principles of Computer Vision (Columbia) is the best visual-intuition companion to Part I in existence: short, beautifully produced lectures on image formation, sensing, and filtering that align almost one-to-one with Chapter 1 through Chapter 4. CMU's 16-385 Computer Vision posts complete slide decks that condense each Part I topic to its load-bearing equations, and the University of Central Florida's CRCV lecture series on YouTube covers the classical pipeline at a comfortable graduate-intro pace.

2. The Classic Papers, Chapter by Chapter Intermediate

Reading the original papers behind Part I is more rewarding than it sounds: the classics are short, concrete, and often clearer than their textbook summaries. Table 8.4.2 lists the primary sources for the algorithms this part used most, with the chapter where each appeared.

Two of those rows deserve a special note: the IPOL editions. The Image Processing On Line journal publishes algorithms with reviewed reference implementations and a browser demo; every claim is reproducible by construction. When you need to understand exactly what a classical algorithm does, including the details the original paper left out, IPOL is frequently the best version of the paper.

from skimage import restoration

doc = restoration.denoise_nl_means.__doc__
refs = doc.split("References")[1]        # numpydoc sections are plain text
print(refs[:290])

----------
.. [1] A. Buades, B. Coll, & J-M. Morel. A non-local algorithm for image
       denoising. In CVPR 2005, Vol. 2, pp. 60-65, IEEE.
       https://doi.org/10.1109/CVPR.2005.38
.. [2] J. Darbon, A. Cunha, T.F. Chan, S. Osher, and G.J. Jensen,
       Fast nonlocal filtering ...

3. Staying Current Without Drowning Intermediate

The research firehose is real: arXiv's computer vision category receives on the order of a hundred submissions every weekday, and image processing work is split across cs.CV and eess.IV. If you can deeply read two papers a week, you are sampling well under one percent of the flow, so the design goal of a reading workflow is the filter, not the feed. A workable funnel has three stages: a broad skim source (arXiv listings or the community-curated Hugging Face Daily Papers), a relevance filter (does it touch your task, with code?), and a depth queue you actually empty. Papers with Code short-circuits much of the work by attaching leaderboards to the benchmarks of Section 8.3, and the CVF Open Access archive provides every CVPR, ICCV, and WACV paper free. Code 8.4.2 automates the first stage of the funnel.

import arxiv   # pip install arxiv

client = arxiv.Client()
search = arxiv.Search(
    query='cat:eess.IV AND abs:"image restoration"',
    max_results=5,
    sort_by=arxiv.SortCriterion.SubmittedDate)

for r in client.results(search):
    print(r.published.date(), "|", r.title[:70])

2026-06-08 | Frequency-Aware Degradation Modeling for Blind Image Restoration
2026-06-05 | A Unified Prior for Joint Denoising and Super-Resolution ...
2026-06-04 | Benchmarking Perceptual Metrics Under Real Sensor Noise ...

A final calibration: classical image processing is a slow-moving field by deep learning standards, which is good news for your reading budget. The bilateral filter is decades old and still ships in every camera; the metrics of Section 8.3 change on a timescale of years, not weeks. Spend your currency budget where the field actually moves fast, which from Part III onward it does, and let Part I's foundations compound quietly. That closes the toolbox, and Part I with it: Chapter 9 begins asking what the images contain.