Back to Projects
Completed Project
DAFEsteel
Steel-surface defect detection with a YOLOv11n model designed for both thin cracks and textured defects.
|
RoleTeam Lead & Model Developer
Timeline2026
TeamSix-person Digilians graduation team
Tech Stack6 Technologies
Mission Brief
For my Digilians graduation project, I led a six-person team building a detector for all six NEU-DET steel-surface defect classes. I designed DAFEGate to combine learned edge features, local texture variance, channel attention, and a residual connection inside YOLOv11n. Across 18 experiments, the model reached 81.98% mAP@0.5 with 2.69 million parameters. I also built a Dockerized FastAPI inference service and published a Gradio demo.
Key Features
Model and evaluation
- Detects six steel-surface defect classes in the NEU-DET benchmark.
- DAFEGate combines edge and texture features with channel attention and residual refinement.
- The final model has 2.69 million parameters and reached 81.98% mAP@0.5.
From model to application
- A Dockerized FastAPI service exposes image inference through a REST endpoint.
- An interactive Gradio demo lets visitors test the detector in a browser.
Engineering Chronicles
Thin cracks and broad surface anomalies need different visual cues.
SolutionDAFEGate uses an edge branch and a local-variance texture branch, then fuses them with channel attention and a residual connection.
Installation Instructions
Source and setup
$git clone https://github.com/hazemelerefey/DAFEsteel.git
Project Access
Technologies
Python
PyTorch
YOLOv11n
FastAPI
Docker
Gradio
Table of Contents
- • Mission Brief
- • Key Features
- • Engineering Chronicles
- • Installation Instructions
