This is a binary classification project built from the ground up. The repo includes the original manually created dataset, the augmented dataset, and code snippets for data augmentation, the used CNN model in tensorflow, model training, and prediction.
This was a mini project I did for a bottle printing company that were interested in automatically detecting double sided vertical seams on glass containers. The soluion was simply to emit a laser beam from one side of the bottle and capture the pattern on the other side while the bottle rotated around its axis of symmetry. The problem then could be formulated as a binary classification between seam and non-seam patterns.
| Seam Laser Pattern | Regular Laser Pattern |
|---|---|
 |
 |
The original dataset consisted of 100 images manually taken and then cropped to 128*128 grayscaled images. Data augmentation was then carried out by flipping and changing the contrast/brightness levels of the images.
The model includes 2 convolutinal layers each having 64 3*3 filters combined with max pooling, followed by 2 fully connected layers. Adam optimizer and binary cross entropy loss were used to train a split of 80% - 20% traing and validation set in 50 epochs.Tensorflow 2 was used to create the model.