yuhaowang2018 / Image-classification-of-ultrahigh-carbon-steel-microstructures

Image classification of ultrahigh carbon steel microstructures

This repo contains two files: "image_crop.py" is used to remove the black bar at the bottom of the micrograph. The jupyter notebook is the main file doing analysis.

In the notebook, the following tasks were performed:

• Preprocess the dataset and split it into train and test parts.
• Extract features from different intermediate layers and compare their performance.
• Use the highest accurate layer to obtain features to train a group of binary SVM classifiers.
• Use voting scheme to obtain a multi-class classifier.

Test error on test set for pairwise two-label classifiers:

	spheroidite network	spheroidite pearlite	spheroidite Widmanstatten	network pearlite	network Widmanstatten	pearlite Widmanstatten
spheroidite	0.00729927	0.0036496	0.13138686
network	0.10714			0.080357	0.0535714
pearlite		0.08333		0		0
Widmanstatten			0.380952		0	0.285714286

Test error on test set for multi-label voting classifiers:

	Multi-label voting classifiers
spheroidite	0.142335766
network	0.107142857
pearlite	0.083333
Widmanstatten	0.380952381

Source of dataset: http://uhcsdb.materials.cmu.edu/