RingClassifier

File Structure

RingClassifier
│   README.md
│   clustered_values.csv   
│   stage1_visualization.html
|
└─── code
│   │   stage1.ipynb
│   │   stage2.ipynb
│   
└─── width0.5_original
|
└─── width0.5_cleaned
|
└─── clustered_imgs
│   │   cluster_1
│   │   cluster_2
│   │   cluster_3
│   │   cluster_4
   

• width0.5_original contains 91 images that are width 0.5. They are original screenshots from the heated coils recordings.
• width0.5_cleaned includes 59 images that are cleaned from width0.5_original. The images where the rings are blended with the background are dropped.
• code contains two python code notebooks for RingClassifier, stage1.ipynb and stage2.ipynb. In stage1, the images are resized to (224, 224) and transformed to grayscale. The pretrained model, ResNetV2, is applied to extract image features. These extracted features are further clustered by K-Means, which assigns a cluster number to each of the images. In stage2, a KNeighbors Classifier is trained using the parameter values (curvature, torsion, and cut degree) and the cluster numbers obtained from stage 1. With this, at inference time, the classifier predicts which cluster the unseen shape most likely belongs to, given its parameter values.
• clustered_values.csv contains five columns: image id, curvature, torsion, degree, and cluster number obtained from stage 1. It is the input data file for stage 2.
• stage1_visualization.html is an interactive 3-D plot of clustered_values.csv
• clustered_imgs includes 4 subfolders, each containing the images grouped into that cluster.

How to run the code

Clone this github repository, run each code block in the ipynb notebooks.