1,019 papers mapped: where vision and deep learning meet wood, and where the gaps are

Introduction

Cameras and image analysis have been creeping into wood processing for four decades: first as simple scanners for grading and defect marking, more recently as deep neural networks that find knots, identify species or guide robots. For a sawmill manager trying to decide where to invest, the literature is hard to navigate. Thousands of papers exist, they use different vocabulary, and it is not obvious which applications are mature and which are still academic exercises.

A 2026 review in the journal Forests set out to map this landscape rather than to compare individual algorithms. Its authors, from the Federal University of Vicosa in Brazil and Universiti Teknologi Malaysia, analysed 1,019 Scopus-indexed publications from 1983 to early 2026 on computer vision, image processing, machine learning and deep learning applied to wood and timber. This article summarises how the mapping was done, what it found, and what it means for a sawmill reader. The title of the original paper describes it as a Scopus-based bibliometric and systematic mapping review.

Research methodology

Search and screening

The authors searched the Scopus database on 9 December 2025, using title, abstract and keyword fields and no restriction on publication year. The query combined two blocks of terms with AND. The first covered vision and learning terms (for example computer vision, machine vision, image processing, deep learning, object detection and image segmentation). The second covered wood terms (wood, timber, lumber, roundwood, wood log, log yard, sawmill, log sorting, wood stack and stack timber).

To be included, a record needed evidence of image-based analysis, an explicit focus on wood or wood products (logs, lumber, engineered wood, panels, or microstructure relevant to performance), a relevant task such as defect detection, grading, optimisation, dimensional measurement, process monitoring or property prediction, and complete basic metadata. Excluded were records with no real image analysis, remote sensing or forest ecology without an industrial processing context, medical imaging unrelated to wood, generic vision or graphics papers without a wood application, records with missing metadata, ambiguous cases, and duplicates. Retracted items were removed. Conference papers, book chapters and short documents were kept.

The numbers tell the story of the funnel: 2,956 records were retrieved initially, roughly 1,890 were removed by automated rules (including a small number of duplicates), leaving 1,066 for manual verification, where 47 more were dropped (45 off-topic and 2 retracted). The final corpus was 1,019 papers. Two expert reviewers checked records and settled disagreements by consensus; where full texts were open access they were used, otherwise title, abstract and keywords. All the most highly cited papers, about 10 percent of the corpus (102 papers), were reviewed manually. Reporting followed the PRISMA 2020 guidelines, and the protocol was registered retrospectively on the Open Science Framework.

Classification of papers

Each paper was assigned a single primary application category out of nine, using weighted keyword scores (positive matches earned points and conflicting terms deducted points, with ties resolved by a fixed priority order; papers with no score defaulted to surface quality control). Separately, each paper was assigned to one of three methodological classes in a hierarchy: Deep Learning if any deep learning term was present, otherwise Classical Machine Learning if any classical ML term was present, otherwise Classical Computer Vision. The authors also built a keyword network and ran community detection on it.

Validation of the automatic labelling

Because the labelling was rule-based, the authors checked it against manual labels on a stratified sample of 97 papers. The method-class labels matched in 78 of 97 cases (80.41 percent accuracy). The application categories matched in 68 of 97 (70.1 percent accuracy, macro F1 of 48.3 percent). This validation is important context for everything below, and we return to it in the limitations.

Key findings

Growth and geography

Research activity has accelerated. Of the 1,019 papers, 99 appeared between 1983 and 1999, 137 in 2000-2009, 134 in 2010-2015 and 649 (63.7 percent) between 2016 and early 2026, averaging about 59 papers a year in that last period. Journal articles made up 61.1 percent of the corpus and conference papers 35.2 percent. The most frequent venues were SPIE Proceedings (41 papers), Forests (29), Forest Products Journal (24), BioResources (22), and then Journal of Forestry Engineering and Wood Science and Technology (21 each). Where the corresponding-author country could be identified (707 records), China led by a wide margin with 283 papers, followed by the United States (45), Sweden (35), France (30) and Japan (28).

What the research is about

Nearly half of the literature is about inspecting surfaces for defects. In contrast, the categories closest to the sawmill floor, log and roundwood processing, dimensional measurement, and sawmill automation and robotics, together account for under ten percent of the papers. The authors identify gaps in dimensional measurement, warp detection, sawing optimisation and shared benchmark datasets.

What methods are used

Classical computer vision (hand-designed image features and rule-based processing) still accounts for 714 papers (70.1 percent of the whole corpus), deep learning for 269 (26.4 percent), and classical machine learning for 36 (3.5 percent). The trend, however, is sharp. Deep learning was absent before 2016 and represented 48.8 percent of papers in 2023 to early 2026. Classical computer vision made up 97.4 percent of papers in 2000-2015 and about 55.8 percent from 2016 onwards. Deep learning is most common in species and origin identification (35.2 percent of that category) and surface quality control (31.9 percent), and rarest in dimensional measurement (9.1 percent).

Citation counts per paper are highest for wood property prediction (16.4 on average) and species identification (15.9), and lowest for sawmill automation (7.6), suggesting that the academic attention does not necessarily follow industrial relevance. The overall average was 11.74 citations per paper (median 4).

Directions the authors point to

The authors name Vision Transformers, multi-modal sensing, edge computing and explainable AI as the promising directions, and call for open benchmark datasets and work on the under-represented industrial tasks.

Practical applications for sawmills

Points that come from the paper

• Surface inspection is by far the best-covered topic, so mills looking for vision-based grading or defect detection will find the most published experience there.
• Deep learning is now used in roughly half of newly published work, but classical computer vision is still widespread, which means that simpler methods remain relevant, particularly for structured tasks.
• Industrially oriented topics (log processing, dimensional measurement, warp detection, sawing optimisation) have comparatively little published work, and the authors list them as research gaps.
• There is a lack of shared benchmark datasets, which makes results from different groups hard to compare.

Portal interpretation (not claims of the paper)

• A mapping review counts papers; it does not measure how well systems perform in production. A large number of papers on a topic does not mean that the technology is commercially mature, and a small number does not mean it is impossible.
• When evaluating a vendor claim, ask what data the model was trained on, whether it was validated on boards from your own species, moisture state and lighting, and what the false-reject rate is. The scarcity of benchmark datasets noted by the authors is a reason for asking.
• The gap in log and dimensional measurement may reflect that commercial scanners in sawmills are developed by equipment suppliers and not always published. This is a speculation of the portal, not a finding of the review.
• For mills with in-house development capacity, the review suggests that pilot projects in surface defect detection have the richest literature to build on.

Limitations

The authors acknowledge several constraints and some others follow from the design. The corpus is Scopus-only, so work indexed elsewhere, as well as non-English or industry reports, may be missed. The classification is based on weighted keywords, not on reading every full text, and its accuracy against manual labels was moderate: 70.1 percent for application categories, with a macro F1 of 48.3 percent. Some categories were labelled poorly in the validation sample, for example dimensional measurement (0 percent accuracy in the 97-paper check), wood property prediction (12.5 percent), sawmill automation (33.3 percent) and log and roundwood processing (28.6 percent), though these categories are also small, so the sample contained few examples. The deep learning label depends on a term being present, and the classical machine learning class was labelled correctly in only a third of validation cases. In this sense the counts for the small, industrially important categories should be read as rough indications.

The review is also a map and not a performance comparison: it does not compare accuracy, speed or cost between studies. Finally, our reading of the paper found a few small internal numerical inconsistencies (for example, the share of classical computer vision is reported as 73.6 percent in the abstract and 70.1 percent in the body, and the number of automated exclusions is given as 1,888 and 1,890 in different places), so exact figures should be checked against the original.

Conclusions

The review shows a research field that has grown strongly since 2016, concentrated on surface inspection and increasingly based on deep learning, with China as the dominant contributor by paper count. It also reveals a mismatch: the tasks closest to everyday sawmill operations, such as measuring dimensions, detecting warp, optimising sawing and automating handling, are the least studied, and open benchmark data are scarce. For sawmill decision makers, the review is best used as a guide to where the evidence is thick and where it is thin, not as a ranking of technologies. Combined with local trials, it can help to frame sensible questions to technology suppliers and research partners.

Original research

• Authors: Gianmarco Goycochea Casas, Zool Hilmi Ismail, Helio Garcia Leite
• Title: Computer Vision, Machine Learning, and Deep Learning for Wood and Timber Products: A Scopus-Based Bibliometric and Systematic Mapping Review (1983-2026, Early Access)
• Journal/publisher: Forests (MDPI), vol. 17, article 112
• Year: 2026
• DOI: 10.3390/f17010112
• Link: https://doi.org/10.3390/f17010112
• Licence: Creative Commons Attribution (CC BY)

This article is an independent summary written by the portal; the original paper should be consulted for full details, tables and the authors' own wording.