Developing a deep learning model for recognizing wood bark in a cross-section image

The aim of this study is to develop and validate a deep learning model architecture based on a convolutional artificial neural network for determining the proportion of bark in a digital image of cross-cut timber. The results were obtained using computer vision and deep learning methods. Reference documentation for the Faster R-CNN, RPN, Mask R-CNN, and Segment Anything artificial neural networks was used in developing the model architecture. The open-source Mask R-CNN and Segment Anything artificial neural network models were used in the software implementation of the model for research purposes (the Mask R-CNN model was integrated into a common stack with the Segment Anything (SAM) model, and the feature map generation functions were transferred to the Encoder block; the memory attention blocks of the SAM network were removed to optimize the solution in terms of the required amount of RAM). The model was trained and validated using the author's dataset of digital images of cross-cut timber, available in the repository. The measured Precision metric values during validation demonstrate that the proposed deep learning model architecture successfully solves the problem of segmentation and bark fraction determination in images. Experiments with the deep learning model allow us to evaluate the relationship between its architecture and the number of trainable parameters with the results of bark fraction determination in assortment images: stabilization of the Precision metric is observed when moving from a model with 38 million weights to a model with 217 million weights. Therefore, further research is warranted to determine the optimal number of trainable parameters that provides the best balance of accuracy and image processing performance.

1. Beliaev N.L. Povyshenie tochnosti tsifrovogo ucheta kruglykh lesomaterialov: Diss. … Cand. Tech. Sci. Petrozavodsk: PetrGU, 2023. 163 p. (In Russ.)

2. The Faster R-CNN model is based on the Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks paper. URL: clck.ru/3PpfEE (accessed February 01, 2025).

3. Han S., Jiang X., Wu Z. An Improved YOLOv5 Algorithm for Wood Defect Detection Based on Attention. IEEE Access, 2023, vol. 11, pp. 71800–71810. DOI: 10.1109/ACCESS.2023.3293864

4. Kim T.K., Hong J., Ryu D., Kim S., Byeon S.Y., Huh W., Kim K., Baek G.H., Kim H.S. Identifying and extracting bark key features of 42 tree species using convolutional neural networks and class activation mapping. Scientific Reports, 2022, vol. 12, art. no. 4772. DOI: 10.1038/s41598-022-08571-9.

5. Kunickaya O., Pomiguev A., Kruchinin I., Storodubtseva T., Voronova A., Levushkin D., Borisov V., Ivanov V. Analysis of modern wood processing techniques in timber terminals. Central European Forestry Journal, 2022, vol. 68(1), pp. 51–59.

6. OPENCV library tools for identifying bark on a cross-sectional image of a tree. – URL: https://clck.ru/3Es7QJ (accessed February 01, 2025).

7. Region Proposal Network (RPN): A Complete Guide. URL: clck.ru/3PpfF3 (accessed February 01, 2025).

8. Robert M., Dallaire P., Giguère Ph. Tree bark rеidentification using a deep- learning feature descriptor. 17th Conference on Computer and Robot Vision (CRV). 2020, pp. 25–32. DOI: 10.1109/CRV50864.2020.00012.

9. Wang R., Chen Y., Liang F., Wang B., Mou X., Zhang G. BPN-YOLO: A Novel Method for Wood Defect Detection Based on YOLOv7. Forests, 2024, vol. 15(7), art. no. 1096. DOI: 10.3390/f15071096.

10. What is Mask R-CNN and how does it work? URL: clck.ru/3PpfGp (accessed February 01, 2025).