Designing DAFEGate for steel defect detection
How our six-person team combined edge and texture features in a compact YOLOv11n detector.
Steel defects do not all look alike. Crazing and scratches are thin, edge-heavy marks; patches and pitting cover broader textured regions. In DAFEsteel, our six-person Digilians team built a detector for all six NEU-DET classes. I designed DAFEGate to give the model separate ways to learn edge and texture cues, then combine them with channel attention and a residual connection.
We ran 18 experiments. The final model reached 81.98% mAP@0.5 with 2.69 million parameters. I also built a Dockerized FastAPI service for image inference and published a Gradio demo so the model could be tested outside a notebook. The repository includes the architecture, evaluation records, and deployment code.