Machine vision for birch boards: 90 percent defect detection, with caveats

Introduction

Grading and cutting sawn timber is still, in many factories, a manual job. A worker looks at each board as it passes, marks knots, cracks and other defects with chalk, and decides where it will be cross-cut. This is slow, tiring and inconsistent, and it limits how much of each board becomes high-value product. As labour costs rise, machine vision and deep learning are being tried as a replacement for the human eye.

The paper "Intelligent Automation Manufacturing for Betula Solid Timber Based on Machine Vision Detection and Optimization Grading System Applied to Building Materials" by Min Ji, Wei Zhang, Xingliang Diao, Guofu Wang and Hu Miao (Research Institute of Wood Industry, Chinese Academy of Forestry) was published in the journal Forests in 2023. It describes a whole production concept for Betula (birch, Betula costata) solid timber used in building materials. The concept chains species identification, log sawing optimisation, longitudinal multi-blade sawing, quality detection and optimised cross-cutting. This summary focuses on the detection and optimisation parts for which the paper gives data, and it flags where the numbers need careful reading.

Research methodology

Imaging hardware

The vision station uses a Basler raL2048-48gm line-scan camera (2048 pixels per line, 51 kHz line rate, 8-bit depth). Illumination is a red stripe light source set at a fixed 60 degree angle between the light and the camera, with adjustable height and brightness. The system is calibrated to about 10 pixels per millimetre, that is 0.1 mm per pixel. Images for species work are 2048 x 2048 pixels, while scanning-station images are 2048 x 12,288 pixels. The processing workstation has an Intel Core i9-12900K, 64 GB RAM and an NVIDIA GeForce RTX 3090 graphics card.

Algorithms

• Species identification: images of bark and stump faces were annotated, augmented and used to train an instance segmentation model. The species covered were birch, spruce, Douglas fir, larch and red pine.
• Pre-processing: histogram equalisation to correct grey-level distortion and improve contrast.
• Defect detection: a convolutional neural network (with pooling and a fully connected layer) and a deep-learning regional object detector to label and frame defects. Camera, image-plane and pixel coordinates are related by a pinhole-camera transformation so that defect size and position can be expressed in millimetres.
• Log sawing: logs are reconstructed in 3D from CT scans (largest log diameter 500 mm), and linear programming is used to maximise yield. Sawing rules depend on the defect type, for example gathering large knots on few boards and sawing parallel to cracks.
• Longitudinal sawing: software simulates sawing positions and combinations and chooses unequal-width sawing to maximise yield. The paper states that a fixed multi-blade saw wastes about 8% on average.
• Cross-cutting: a two-objective model, considering total billet length (yield) and total billet value (economics), with the final plan chosen by maximum yield. Cross-cut sizes range from 1200 to 6300 mm in length, 30 to 150 mm in width and 12 to 100 mm in thickness. Detection speed can be set to 150 m/min or 214 m/min.

Material and test set-up

The birch came from Latvia and was kiln-dried to 12 to 15% moisture content (70 to 80 degrees C, 60 to 70% relative humidity), with a density of 0.5 to 0.7 g/cm3. A total of 4,000 randomly selected boards were arranged in 10 packages of 400 boards. A typical board was about 4000 mm long, 58 mm wide and 43 mm thick. Tests were done at a factory in Shaowu, Fujian, China, over the period 2021 to 2023. The defect classes evaluated were bark, dead knot, crack, heartwood, hole, inbark, mineral and sound knot, as well as geometric faults such as material lack, wane, under-thickness and under-width.

Key findings

Detection accuracy

The authors report these headline accuracies:

Among defect classes, heartwood was recognised best (97.97%), followed by inbark (93.12%). Bark, dead knot, crack and heartwood were all above 85%. Holes and mineral streaks gave the lowest accuracy. For size and position, the paper cites, for example, crack length at 93.98%, sound knot length at 92.98%, crack width at 91.70%, and Y-axis position of cracks at 94.80%.

Dimensional measurement error was small: 0.31% for length, 1.16% for width and 1.18% for thickness.

Yield

Most boards fell in the 81 to 90% yield band, followed by 71 to 80% and 90 to 100%. The batch used for the economic comparison had a volume of 33.56 m3.

Speed and throughput

Scanning took 0.311 s per board and image analysis 0.368 s per board. The authors state that the average decision speed is at least 60 m/min against 20 m/min for manual marking, and from this they derive an annual capacity of more than 6,451,200 m, equal to 16,089.29 m3 of additional throughput.

Economics as reported

For the test batch, the optimised value was reported as USD 12,844.52 (CNY 88,241.89) versus USD 11,236.27 (CNY 77,193.19) for manual marking, a difference of USD 1,608.25, which the paper presents as a 14.13% improvement in comprehensive utilisation value. The paper also states a saving of 3 to 4 workers (about USD 34,758 per year) and gives projections for annual gains at assumed monthly volumes.

Practical applications for sawmills

Points taken from the paper

• A line-scan camera with controlled stripe lighting, calibrated to 0.1 mm per pixel, can inspect boards at line speeds of 150 to 214 m/min, with a per-board scan and analysis time of well under one second.
• Defect detection results are then passed to an optimiser that proposes cross-cut plans, replacing the manual marking step.
• The authors argue that combining CT-based log sawing optimisation, flexible-width longitudinal sawing and defect-aware cross-cutting can raise utilisation compared with fixed multi-blade sawing, for which they cite an average waste of about 8%.

Portal interpretation (not claims made by the authors)

• Hardwood and joinery mills that currently rely on manual marking may find this paper a useful benchmark for expected detection accuracy: roughly 90% for type and size, with weaker performance on holes and mineral streaks. Those weaker classes are often the ones that determine grade in appearance products, so a mill should test them on its own material.
• The economic figures should be treated as indicative only. A single batch of 33.56 m3 is a small sample, and the paper's own numbers contain inconsistencies (see below). A mill should run its own pilot with its own price list before relying on the percentages.
• The speed advantage over manual marking (60 versus 20 m/min) matters mostly if the downstream equipment can use the extra throughput. The whole line, not only the scanner, has to be balanced.
• Training data matter. Any vision grading system will need images of the mill's own species, drying quality and typical defects.

Limitations

The authors acknowledge that detection errors remain and that work to reduce them is ongoing. Several limitations are visible in the paper:

• Scope: the timber is birch from a single source (Latvia) tested at a single factory, so generalisation to other species, regions or surface conditions is untested.
• Misclassification: holes are sometimes read as knots, dead and sound knots are confused, stains are taken for minerals, heartwood can look like sapwood, and light bark can be mistaken for mineral streaks. Some location estimates gave negative values, especially along the Y-axis for under-width defects, and outliers were replaced by means in several data sets.
• Internal inconsistencies in the numbers, which we noticed while reading: the abstract and conclusions map the percentages for type, size, location and yield in a way that is ambiguous; USD 1,608.25 divided by USD 11,236.27 is about 14.31%, not 14.13%; the listed volume terms sum to about 33.62 m3, not 33.56 m3; and the production assumptions differ between sections (12,000 m3 per month in one place, 1,000 m3 per month in another). Currency conversions use fixed rates that are not explained.
• The CT-based log sawing optimisation is described, but no quantitative result for it is reported, so its benefit is not demonstrated in this paper.
• We read the open-access full text, but a few thousand characters at the very end of the page were not retrieved, so minor details there are not reflected.

Conclusions

The paper presents an integrated vision-and-optimisation concept for birch solid timber and reports detection accuracies near 90% for defect type and size and about 92% for location, with dimensional errors around 1%. It argues that this can replace manual marking, cut labour and raise the utilisation value of sawn timber. The direction is plausible and consistent with wider industry trends, but the economic claims rest on one batch, one species and one factory, and the paper contains numerical inconsistencies. Sawmills should read it as a promising case study and design their own validation before investing.

Original research

• Authors: Min Ji, Wei Zhang, Xingliang Diao, Guofu Wang, Hu Miao
• Title: Intelligent Automation Manufacturing for Betula Solid Timber Based on Machine Vision Detection and Optimization Grading System Applied to Building Materials
• Journal: Forests, 2023, volume 14, issue 7, article 1510
• DOI: 10.3390/f14071510
• Link: https://www.mdpi.com/1999-4907/14/7/1510
• Licence: Creative Commons Attribution (CC BY), open access

This article is an independent summary by the portal; the original paper should be consulted for full details.