A deep convolutional transcriber

As a first attempt at a model which might do the transcription, I’m using a straightforward deep convolutional approach: Six 2x2 strided convolutional layers to reduce the dimension of the image, a single fully connected layer to extract the transcribed numbers from the convolutional features, and a softmax layer to express the predicted transcriptions as probabilities for each digit. With ELU activations, and dropout in each layer to resist overfitting.

Schematic of the deep-convolutional transcriber model

• Functions to provide training/test Dataset (common across models)
• Model specification
• Model fit script

This model is not in any way optimised to match the problem - it’s a first attempt using standard methods. It’s not perfect, but, after 200 epochs of training, it has considerable skill:

• Example image with transcriptions
• Transcription accuracy by digit
• Transcription accuracy by page location