TechnologyVisual Intelligence
From a single image to material intelligence.
How a visual observation can become the first structured context of a stream, without pretending to be a measurement.
CENTRAL QUESTIONWhat can we learn from an image, and what can’t we?
AN IMAGE CAN SUGGEST
- Likely material or family
- Form and presentation
- Apparent color
- Visible contamination
- What is missing to confirm
AN IMAGE CANNOT ESTABLISH
- Actual purity
- Chemical composition
- Moisture
- Grade or specification
- Exact value
01CONTEXT
A lot of information about a material is visible before it is measured: its shape, its color, its presentation, the obvious contamination. That information usually stays in the head of whoever is looking, or in a photo sent over a messaging app. The problem is not that information is missing; it is that it does not become a record that can be compared with others.
Computer vision looks like an obvious answer, and it is half of one. Recognizing that something looks like a bale of PET is one thing; stating its purity or grade from the same image is another, and confusing the two is the most expensive mistake.
02OUR HYPOTHESIS
An image is an observation, not a measurement. If the system treats it as an observation, it can be very useful: it produces a first structured reading, declares what it cannot know and proposes how to confirm it. If it treats it as a measurement, it produces a convincing and wrong number.
Waste Scanner uses vision and multimodal analysis with that discipline. Each field is presented as a visual estimate, and missing information is flagged explicitly. For that very reason we do not show confidence percentages or promises of accuracy.
An honest reading of an image is more useful than a confident one.
03HOW IT WORKS
The path from an image to a record has six steps. The last two are the ones that distinguish an observation from a claim: flagging what evidence is missing and saying what the next validation would be.
- 01IMAGEOne or more photographs.
- 02OBSERVATIONSWhat can be seen.
- 03MATERIAL HYPOTHESISWhat it might be.
- 04STRUCTURED FIELDSForm, color, presentation.
- 05MISSING EVIDENCEWhat cannot be known.
- 06NEXT VALIDATIONHow to confirm it.
04EXAMPLES Illustrative example
Five different materials show how what an image contributes is split from what remains pending. In none of them does the image replace a measurement.
| Material | What is observed | What is left to confirm |
|---|---|---|
| PET (bale) | Presentation, color, visible labels | Composition, moisture, internal contamination |
| PCB boards | Board type, visible components | Origin of the rejection, composition |
| HDPE (purge) | Shape of the purge, color | Grade, contamination |
| Metal (offcut) | Shape, surface appearance | Alloy, composition |
| Corrugated cardboard | Presentation, apparent moisture | Contamination, volume |
05CURRENT BOUNDARY
The visual reading is the starting point of a record, not its conclusion. Today it produces preliminary fields and questions for whoever knows the material; it does not produce a verified specification. That difference is developed in Material Specification.
06WHAT WE’RE LEARNING
The hardest part is the information an image cannot give: moisture, composition, grade. We also learned that the order of capture matters. Asking for a second photo or an extra data point usually contributes more than trying to squeeze a first reading, and showing what is missing is part of the result, not an apology.
07WHAT COMES NEXT
More materials, validation against measurements and laboratory results, and a more direct connection between the first reading and the identity of the material.