Few-Shot Computer Vision for Industrial Wood Grading

Background and industry context

A leading European timber processor needed to automate quality grading across thousands of logs annually. High-grade logs command premium furniture prices while lower grades serve construction or pulp markets. The core challenge: accurately detecting logs in real-time conveyor belt imagery regardless of size, shape, or condition, creating the foundation for automated classification that determines optimal use from premium furniture stock to industrial pulp.

Result

Delivered a production-ready detection module achieving 99%+ accuracy with just 50 manually-annotated training examples. The system processes logs at full conveyor speed, enabling real-time quality decisions. This module paved the way for downstream applications in clustering, and classification.

Example of wood face detections (source)

Description of the solution

Implemented Eager Few-Shot Object Detection based on RetinaNet, starting with ImageNet pre-trained weights and fine-tuning on 50 hand-drawn bounding boxes. The model worked well to draw bounding boxes around logs of varying sizes, shapes, and conditions (straight, healthy timber and crooked/damaged logs). For quality analysis, I fed the detected log regions through t-SNE/UMAP clustering algorithms which separated logs into groups corresponding to size categories (large vs. small) and wood health (healthy vs. rotten), validating that the visual features captured meaningful quality indicators.

I developed the core detection module while collaborating with senior data scientists on adjacent system components. Responsible for algorithm selection and implementation strategy, curation of training dataset, model training and optimization, and performance validation against production requirements.

Technologies used

• Python & TensorFlow – Deep learning model development and training
• OpenCV – Image preprocessing and augmentation
• RetinaNet – Object detection backbone architecture
• Eager FSL – Few-shot learning framework for minimal training data
• t-SNE/UMAP – Dimensionality reduction for unsupervised quality clustering
• NumPy/Pandas – Data manipulation and analysis