Explainable AI for Visual Defect Inspection

Graduate Dissertation

Defect highlight defect highlight


  • Developed and studied XAI algorithms that generates saliency maps according to the importance of each corresponding pixels of the input test image towards the Machine Learning model’s predictive accuracy, with the aim of decoding complex black-box models.

  • This repo contains keras implementation of few XAI algorithms on NEU surface defect dataset.

Directory Structure:

  • NEU steel surface defect database
    • Original train images
  • NEU steel surface defect database - Test Split
    • contains 180 test images (with 10% partition)
  • Finetune_on_NEU_dataset.ipynb
    • Training a MobileNet model on the NEU surface defect dataset using Transfer Learning.
    • Visualizing the important features of test images using model agnostic LIME and SHAP algorithms.

Sample images from NEU dataset : Sample images from NEU dataset

Mahesh Sudhakar
Mahesh Sudhakar
Computer Vision Research Engineer

Computer Vision | Robotics | Machine Learning