Jindal Stainless · Stainless Spark · Problem statement 1
Surface defects start early, at the caster, the descaler or the furnace, but a coil is graded only at the end of the line. At 300 m/min no inspector can see the whole surface.
This page is the detector. Click a sample, or click the box to upload your own photo of steel: it finds each defect, names it, rates how serious it is and says what to do about it. Everything runs on your own machine. Nothing is uploaded.
Click a sample frame
Each defect it found, how sure it is, how serious it is, and what the mill should do about it. Click a row to highlight that defect on the image.
Nothing analysed yet. Pick a sample above, or drop an image.
Likely cause and recommended action appear here, one card per defect type.
A neural network's raw score is not a probability. We corrected it using images the model was never trained on, so the confidence you read is honest. After the correction, the stated confidence is within about 3 percentage points of how often the detector is actually right.
Each defect type has a base level: an inclusion, for example, is critical because it cannot be rolled or pickled out, while a light scratch is medium. The detector raises or lowers that level with its confidence and with the size of the defect, and reports low, medium, high or critical. The causes and actions come from published steelmaking practice.
Measured on images held back from training. Nothing on this page was tuned to them.
It is trained on real clean strip as well as on defects, so an alarm means something: it finds 9 in 10 real defects while fewer than 1 in 20 clean images raise one.
| On images it never saw | Result | What it means |
|---|---|---|
| Defects caught on real mill strip | 258 of 288 | 9 in 10 defects found |
| Clean mill strip wrongly flagged | 14 of 300 | fewer than 1 in 20 |
| Right defect type named (lab images) | 178 of 180 | 99% named correctly |
| Confidence against reality | within ±3 points | a "70%" alarm is right about 70% of the time |
| Speed in a laptop browser | ~55 ms per image | no GPU, nothing uploaded |
Test data: 180 lab photographs of six defect types, and 588 images of real production strip from a steel mill (288 with defects, 300 clean). Public research datasets; no Jindal Stainless data was used.
Four steps, all inside this page. On a mill line the same model reads camera frames instead of uploads.
An image of the strip surface comes in. Wide images are split into overlapping square windows so small defects keep their detail.
A compact neural network marks each defect and names its type. It was trained on lab images and on real clean and defective mill strip.
Each defect gets an honest confidence and a severity from low to critical, based on its type, the detector's certainty and its size.
The likely cause and the recommended action are shown for the process that caused it, from melting and casting to handling.
What a pilot on a real line would need to prove.
Stainless surfaces and finishes look different from hot-rolled carbon steel. Performance on a stainless line has to be measured there, after training on that line's own clean and defective strip.
The detector knows these are not clean steel, and says so, but claims no root cause. In a pilot they would be mapped to the plant's own defect catalogue.
Public datasets have few or no examples of these. Covering them needs targeted images from the line, and edge cameras for edge cracks.
On a line the cameras only ever see strip, and a separate check that rejects non-steel images is part of the full system. This demo page does not include it.