2026/05/13 by Prabu Ravindran, Flavio Ruffinatto, Alberta Asi Ebeheakey +3 · 1 voice
Chemistry · Engineering · Environmental Science · #Wood and Agarwood Research #Wood Treatment and Properties #Remote Sensing and LiDAR Applications
paper · doi:10.1163/22941932-bja10216
openalex publication_date 2026/05/13 · openalex created_date 2026/05/15 · openalex updated_date 2026/07/23
Summary Macroscopic digital images suitable for computer-vision wood identification typically present a field of view and level of anatomical detail sufficient for human-based macroscopic identification. While humans trained in wood anatomy can recognize, locate, delineate, and measure anatomical features in such macroscopic images, image-level classification algorithms cannot. Determining which pixels in an image, specifically, correspond to growth ring boundaries (GRB) could provide additional explicit quantitative data for analysis. We present a conceptual typification of twelve visual cues in digital images that delimit GRB as a step toward bridging the gap between human-defined anatomical features and the digital image characteristics representing those features. To detect GRB using these cues, we trained a multiscale deep supervision edge detection model on a dataset of 823 XyloTron images from 40 softwood species representing 9 genera and developed a simple post-processing method to convert the predicted GRB probability image to a binary mask. The model predicted all 381 GRB in the test set images, and 31 additional false-positive GRB. Once the false positives were removed, a comparison of the average differences between the ground-truth and predicted GRB showed that 73.5% were within 1 pixel while 98.2% were within 3 pixels (approx. 9.3 μm or less).