Steel surface defect classifier (NEU-CLS)

ResNet-18 fine-tuned on the six NEU-CLS defect classes, running in your browser with ONNX Runtime Web. Model: senanurcetin/neu-steel-defect-resnet18 · project: visual-qc-project.

360/360 on NEU-CLS under both a random and a blocked split (the project's handcrafted-feature Random Forest: 0.939). A saturated benchmark: accuracy falls to roughly 55–65% with slight blur and 51–59% at quarter resolution (robustness test), so validate on your own line before trusting it.