Visual Defect Detection for Luxury Manufacturing, Kasra Zamanian

Setup

This was my master's thesis at Bocconi. A top luxury house wanted to automate defect detection on leather hides, a job that took an expert operator 10 to 20 minutes per hide and could only be done by experts.

The starting point was, realistically, almost nothing:

no historical data and no images,
no rig or machinery to capture data,
no clear idea of how to approach the problem,
and hides that were nothing like one another.

What they needed was closer to a magic wand than an engineering project. On top of that they wanted close to real-time detection, accurate localization, and a model that transferred easily to new hides. The one mercy: they could tolerate false positives, but not false negatives. Those constraints ruled almost everything out and pointed squarely at YOLO.

I had already done defect detection at my startup (AIWAY) and knew vision models well, so my first thought was the naive one: easy. Collect data, annotate, train a YOLO, done.

Why it was hard

That confidence died the moment I saw the leather. The operator laid a 2m by 1m hide on the table, and I genuinely thought it was a clean sample: I could not see a single flaw. He said it was a dirty one, maybe 70 to 100 defects. I thought he was joking.

A hide the operator graded as 'dirty', with dozens of defects on it. Can you spot even one? It took me three months.

He pointed at what looked like empty leather and I saw nothing. Then he tilted a flashlight at a precise angle and, for a second, a faint abnormal line appeared: a vein. Then a black speck I would have called dust: a "dot", a defect smaller than the point of a needle.

The reveal. The blue box (red X) is just dirt. The green boxes, a vein and a stretch mark, are the actual defects.

That was the humbling moment. Forget YOLO. Before I could teach a model to see these defects, I had to teach myself.

The capture rig we settled on: a matrix camera with four corner lights, tuned over hundreds of trials.

Tiled inference: the trick the industry missed

What happened: How does a model find a defect smaller than a millimeter? You shoot it close up. But then big defects fill the frame with no background left for context.

The fork: Optimize the imaging for small defects or for big ones. Either choice broke the other.

My call: Shoot from a moderate distance at very high resolution, so big defects keep their context and tiny dots still get enough pixels. But you cannot feed a 6k image to a model: it is slow, and a few-pixel dot has no weight among millions. The answer was tiled inference: slice the 6k image into overlapping 640px tiles (enough context for big defects, more relative weight for small ones), train on the tiles, and at inference run on the tiles and stitch the predictions back onto the full image.

What it led to: The imaging and tiling were the foundation. But the model still was not good enough, and fixing that taught me the most.

From 0.79 to 0.91: teach 'normal', then do error analysis

What happened: Even with tiling and transfer learning, the results were poor: an explosion of false positives, tens of phantom detections per frame. We had focused so hard on defects that we never taught the model what normal leather looks like. Adding background images with no annotations killed the overfitting and got us to 0.79 precision and 0.67 recall. Good, not great, and I was stuck.

The fork: Keep tweaking architecture and hyperparameters blindly, or find out why the model was failing.

My call: I was reading Andrew Ng's Machine Learning Yearning on the side, and his chapter on error analysis was the unlock. Two of us went through 100 predictions one by one against the ground truth, categorized every error in a spreadsheet (missed defect, misclassification, false positive), noted the likely cause, and ranked the categories by frequency and impact.

What it led to: The data was blunt. Most errors came from bad bounding-box cuts, an artifact of the tiling step, and a second cluster had "no clear reason", which we traced to undertraining (too few epochs). Fixing those took us to 0.91 precision and 0.72 recall across every defect category.

Impact

The thesis was a hit: 0.91 precision and 0.72 recall on defects a layperson cannot even see. The results were strong enough that, a year later, the company came back: the client wanted a standalone product built around it. I took the offer and led that build, which became its own story: Automated Leather Inspection Machine.

Takeaways

You cannot automate a process you do not understand. Months of learning to find the defects by hand was the real unlock, not the model.
Often the breakthrough is a simple reframing, not a fancier model: tiling for resolution, and background images to teach the model what "normal" looks like.
Error analysis beats guessing. Reading 100 predictions by hand told us exactly what to fix when the metrics alone could not.