This file is a step by step tutorial to practice basics of transfer learning on an industrial dataset composed of pictures, for a use case of quality control. You will also find the dataset and Jupyter notebooks with code that can help you address the questions.

• available in casting_512x512.zip or on Kaggle: https://www.kaggle.com/datasets/ravirajsinh45/real-life-industrial-dataset-of-casting-product
• This dataset provides image data of impellers for submersible pumps.
• There are many types of defect in casting like blow holes, pinholes, burr, shrinkage defects, mould material defects, pouring metal defects, metallurgical defects, etc.
• Raw dataset = 519 ok + 781 defect
• (Kaggle also has an already augmented dataset = 7348 pictures)

• Use this Keras tutorial:
  • https://keras.io/guides/transfer_learning/
  • (the tutorial is mirrored in 231122_keras-transfer_files.zip)
• Take a pretrained DL image classification model
• Freeze weights
• Remove last layer
• Replace it by what makes sense to your question
• Train only the part you added
• See 01_defects_transfer.ipynb for example code

• With different kinds of image classification models
• With different layers on top of the pre-trained model
• With or without data augmentation
• With a different number of input pictures

Possible output:

• You can look into explainability for computer vision algorithms

• SIFT (or AKAZE)
• Bag of words
• Classification with SVM
• See 02_defects_AKAZE_SVM.ipynb

Used in 2022 and 2023 with AI engineering students in Bordeaux.