Steel-surface defect detection for an inline inspection camera

Rheinmark rolls thin steel strip for automotive suppliers at two plants, one in Germany and one in Pune. At the end of each line, two inspectors watch strip pass at 3 metres a second and mark surface defects by hand. They miss things. Last year 14 coils were rejected by a customer for scratches and inclusions that went through inspection, at an average cost of €9,200 per rejected coil, plus freight.

The plant has installed a line-scan camera that cuts the strip image into 200 × 200 pixel tiles. The quality manager wants a model that labels each tile as one of six defect types or clean, so inspectors only look at flagged tiles. Her targets are concrete: - recall of at least 95% on inclusions and scratches, the two defects customers reject coils for; - no more than 1 false flag per 50 clean tiles, or inspectors will start ignoring the alerts; - under 20 ms per tile on the plant's CPU-only edge PC, so the model keeps up with the line.

You will use the public NEU surface defect dataset (1,800 greyscale images, six defect classes, 300 each), which matches the camera's tile size, and generate clean tiles by augmentation or from a second public set. You will build a simple baseline, fine-tune a small pretrained CNN, analyse its errors class by class, tune decision thresholds per class and measure CPU latency.

Done looks like a training repo, a confusion-matrix report with per-class recall, a latency benchmark, and a short note for the quality manager on whether the model is ready for a shadow trial on one line.