Navatej1/Automated_Defect_Detection

Casting Defect Inspector

This Space classifies controlled top-view images of submersible-pump impellers as Defective or OK and displays an optional Grad-CAM explanation.

Required files

The training notebook automatically generates this complete folder. The repository root must contain:

• —app.py
• —requirements.txt
• —casting_defect_model.keras
• —model_metadata.json
• —optional examples/

The binary output is interpreted as probability of a defective casting. The decision threshold is selected using validation data and stored in model_metadata.json.

Limitations

This is a research and decision-support prototype. It is not validated as a production quality-control system, does not identify defect type, and is intended for images captured under conditions similar to the training data.