An image-processing problem can begin before the software sees the image. In multi-reel X-ray counting, the beam and sample arrangement affect whether separate components remain distinguishable. More training data cannot be assumed to recover information that the projection has obscured.
One reel and four reels create different views
In the TruView PC2 development Fireside Chat, Griffin Lemaster describes moving from a centered reel to four reels positioned across the field. The wider arrangement introduced overlap and uneven parallax around the individual reels.
With a diverging beam, features away from the central axis are viewed at different angles. The counting task therefore depends on the acquired projection as well as the component-recognition algorithm.
Training did not remove the geometry problem
The team gathered reel images and spent substantial effort labeling parts. Lemaster describes recognizing that additional training was not the complete answer to the uneven projection conditions. The development task had to address the source of the image variation.
The conversation then explains changing how the reels align with the source’s beam direction. The objective was a more manageable geometry around each reel, not a claim to eliminate the physics of a diverging beam.
Co-design handling and image analysis
The engineering lesson is to examine whether sample positioning can make the required features more consistently observable before adding algorithm complexity. Mechanical arrangement and software are parts of the same inspection method.
Evaluate the actual reel population
This is a development account, not a current accuracy or cycle-time specification. Different components, reel arrangements and count requirements should be evaluated with representative samples. Discuss your parts-counting application to connect the image geometry to the decision the system must make.