When scanners learn from each other

AI is changing how scanners in sawmills work together. At MiCROTEC, data from different scanners is used to make each one smarter, improving accuracy and performance across the entire production line. We spoke with Thomas Gagliardi, Product Manager for log solutions, about how this idea works in practice, the training methods behind it, and where AI in wood scanning is headed next.

MiCROTEC provides an extensive portfolio of solutions for scanning and managing material processed within a sawmill. Considering that each scanner is designed to meet the specific requirements of a given step in the production process, it follows that every scanner is built upon hardware and software technologies that differ from the outset. This means that each measurement instrument inherently has its own technical capabilities as well as certain intrinsic limitations.

With the advent of Artificial Intelligence, it has become possible to enhance the capabilities of a scanner through the software that processes the raw data it provides. MiCROTEC’s approach is based on the concept that, although each scanner considered individually presents certain limitations, when scanners are interconnected within their operational context, the different technological types can effectively compensate for one another. Moreover, because each scanner is integrated into different stages of the production process, the information derived from material as it becomes progressively more processed can serve as valuable ground truth for training AI models dedicated to each scanner.

As an example, consider a log scanner that measures the entire object, and a board scanner that inspects material originating from what was once a single, monolithic object prior to cutting. Sharing this information across the various production stages—combined with the different hardware sensor technologies of each scanner—constitutes an expanded and essential source of knowledge for providing AI models with the most complete and accurate training possible.

As mentioned, the key factor underlying the reliability of an AI model is the quality of its training set. An AI model must be viewed as a system capable of learning to solve a problem even when presented with scenarios that differ from the norm. This concept captures the potential of an AI‑based solution: greater flexibility compared to a traditional iterative algorithm designed to solve a problem in a predetermined sequence of steps. That said, an AI model becomes more reliable as the training set becomes broader, more complete, and more accurate.

MiCROTEC implements two different approaches in this regard: training with CT data and cross‑training between scanners.

The first approach uses data from tomographic scanners—such as the MiCROTEC CT Log or the BIOMETiC Mito—to provide a comprehensive, extensive, and accurate ground truth for AI models developed for other scanners, without requiring a CT scanner to be pre-installed in the production line. Consider, for example, the measurement of the under‑bark shape of logs using a scanner equipped with two static X‑ray sources (radiography), such as the MiCROTEC Logeye 302 or similar state‑of‑the‑art systems on the market. Although radiography can see beneath the bark, it provides only partial information because the X‑ray projection condenses all traversed structures into a 2D plane. In practical terms, this type of scanner delivers direct measurements at only 4 of the 360 points constituting the circular surface of a log.

In contrast, the CT Log scanner measures the full 3D under‑bark geometry of each log thanks to its tomographic technology. With this precise reference and the ability to automatically generate a vast amount of data, a multi‑sensor approach has been implemented in the Logeye 302 scanner—leveraging all its sensors (laser, high‑resolution color cameras, and X‑ray units)—and AI models trained with CT Log data now enable unprecedented accuracy in estimating the under‑bark shape of logs. Likewise, using the extensive and accurate dataset provided by the CT Log, the performance of the Logeye 302 has been significantly enhanced with regard to identifying knot characteristics (sound versus dead), determining pruning depth (pruned/unpruned), accurately estimating growth ring width inside the log, and detecting top break and compression wood.

Another example is the use of the BiOMETIC MiTO scanner, the tomographic scanner employed in the food industry. Due to its smaller dimensions relative to the larger CT Log, it is suitable for scanning boards. This has allowed the precise and complete tomographic data to be used for training the AI models installed on Goldeneye scanners, specifically for the Pith Finder for Knot Connection function. This approach aims to provide high‑quality AI models to many MiCROTEC scanners before they are even installed in a sawmill’s production line.

MiCROTEC’s second approach is cross‑learning between in‑line scanners. MiCROTEC introduced its MiCROTEC Connect solution a few years ago; it provides a precise traceability system for each piece of wood throughout the entire production process—from the raw log to the final board ready for sale. With this approach, it becomes possible to link each processed piece of wood to its earlier stages from any scanner present along the production line.

For instance, given a board in the Green Mill scanned by the Goldeneye Transverse, one can determine from which log it originated and, more importantly, its exact position within the log prior to cutting. By leveraging MiCROTEC Connect and interlinking the Goldeneye Transverse scanner in the Green Mill with the CT Log in the Logyard—together with a Logeye Fingerprint scanner and a Truespin rotation control unit at the sawline infeed, plus the Logeye Cant scanner between the first and second sawing breakdown—the detection of sapwood in the Goldeneye Transverse has been greatly improved. For many softwood species, sapwood and heartwood are difficult to distinguish using visual images alone. However, due to density differences, this task is straightforward for the CT Log. Using virtual boards generated by the CT Log and their heartwood/sapwood classification as in‑line ground truth, this improvement was successfully implemented in the Goldeneye Transverse.

This cross‑learning approach takes place in real time during a sawmill’s normal production and can deliver continuous day‑to‑day performance improvements.

MiCROTEC is a leader in innovation, and therefore, the horizons ahead of us are numerous. Consider, for instance, the solutions already in place to improve the accuracy of diameter measurement during sorting when logs are handled with bark in the Logyard, only to be debarked later at the Sawline Infeed before cutting—an extremely critical step for sawmill optimization. Then, reflect on the many possibilities offered by high‑definition images from MiCROTEC color cameras, now installed on nearly all MiCROTEC scanners.

What remains fundamental—and central to our research—are the needs of the global sawmill market. Technology must serve this industrial sector and be capable of addressing the demands of individual markets, each with its own wood species, production characteristics, and specific end-products. The overall trend in the industry points toward greater automation, providing solutions capable of achieving performance levels comparable to human operators, but with far greater consistency.

Moreover, the integration of AI‑based solutions with the possibility of training on CT data or via cross‑learning keeps the door open to future developments, as these methods are flexible and scalable in response to new requirements emerging from the global wood market in the face of ongoing environmental and economic challenges.