Classifying "QR codes"

Part 1 (2019) Advanced (Part 1 v3)
1

Hi all,

I’m interested in training a CNN classifier on non-visual data that can be represented as binary 0/1 matrices (similar to QR codes). Each data point is a NxN matrix (1 channel, so 1xNxN), and around 5-10 possible labels (one label per matrix).

Presumably the “usual” pre-trained models aren’t well suited, here (again, think QR codes - “meaningful” lower level features, if exist, will be be entirely different). However, I’m still interested in using the fastai library for its many useful functionalities and defaults. If anyone has any experience with something similar, I’d be very interested in hearing tips and advice on this.

Thanks!

1 Like
2

Hi @adamh , you may want to take a look at the Time Series/ Sequential data study group. There you will be able to see examples of how CNNs can be successfully applied to 2D transformations of univariate time series that actually look like QR codes.

1 Like