¶ Automated sweep measurement project
Description: This study aims to support the development of an automated system for measuring
log curvature (‘sweep’) using computer vision. Detection of log sweep has relied on manual
measurement which is very inefficient, or the experience of machine operators. To date there
is no accurate and fast way to determine the sweep of logs, leading to potential
misclassification and reduced value recovery at time of harvest.
This project shows how computer vision and deep learning can build a model that
automatically identifies and calculates log sweep. The model was built based on the deep
learning framework of PyTorch and the deep learning environment of MMDetection.
Data was collected and pre-processed to train the model. A total of 100 images were
manually annotated using LabelMe software. The annotated JSON files were programmatically
transformed into coco-type files for deep learning. The pre-processed data was divided into a
training set, a validation set, and a test set in a ratio of 4:1:1. This ratio avoids overfitting of
training results.
The image parameters were modified based on the existing Mask R-CNN base model
and training was started to obtain a model that only detects the stems in the image. The
coordinates of all pixels of the mask can be extracted from the detection results. Using the
coordinates that make up the shape of the log, a simple mathematical model can then
calculate sweep.
Most of images evaluated were more than 90% accurate in identifying the stem, which
was the desired goal. The programme currently calculates a relative sweep, which is the ratio
of the offset of the centre-line to straight line as drawn from the two end mid-points. In future
work, accuracy can be improved by, for example, adjusting the brightness or contrast of the
image and enhancing the existing model. By interfacing with the harvester head that captures
log length and small end diameter, the sweep can be referenced against the log sweep
specifications to see if it adheres to the expected standard. The future goal is to achieve real-time, fully automated video detection and sweep calculation for harvesting and processing
operations.
¶ Details
| Company | Country | Type | Website | Available in Australia? | Year | Status | Remarks | TRL | Employees | Colour | Inventors | Patent Model. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Interpine | New Zealand | Link |