Self-learning flaw detection using deep learning

Background

Where it comes from?

Machine Vision

  • Bullet One
  • Bullet Two
  • Bullet Three

Inspection

  • Bullet One
  • Bullet Two
  • Bullet Three

Datasets

Cutting Heads

  • 301 images of intact heads
  • 276 images of broken and damaged head

http://pitia.unileon.es/varp/node/439

Chocolates

  • 40 images of intact chocolates
  • 10 images of defective chocolates

http://pitia.unileon.es/varp/node/439

Textiles

  • Images of cloths with no defects
  • Images of cloths with different types of damages
    • Holes
    • Stains
    • Cuts

(To be created)

Methodology 

Techniques and Framework

Transfer Learning

Technique that allows us to reduce training time and dataset size.

  • Use existing architectures created by top companies
  • Fine tune it to suit your needs
  • Reduce modelling time

http://cs231n.github.io/transfer-learning/

TensorFlow

https://www.tensorflow.org/how_tos/image_retraining/

  • Deep Learning Framework
  • Tutorials and examples
  • Community maintained
Made with Slides.com