omarbakker / digibill

digibill

TODO

Line/Word Detection

• MSER segmentation
• Kmeans clustering
• Kmeans++ clustering
• Hierarchical Clustering
• Stopping criteria for hierarchical clustering
• DBSCAN (most likely to yield reasonable results)
• OPTICS
• findCountours from openCV (see dzone article)

Dataset Generation

• Generate images of text
• Generate images of text with a given font (color, size, type, bold/italics/underline)
• Find an appropriate list of fonts to use
• Find an appropriate list of transformations to make data realistic
• Find a corpus (UPC database)
• Generate the synthetic dataset using the above steps

Design/write the training procedure

• CNN to extract features from word images
• Bi-LSTM takes sequence of CNN extracted features
• CTC classifies
• Batch norm where appropriate

Important links

• Dropbox Article
• Convert TF to Coreml
• OpenCV
• Dropbox apply
• Dzone article
• CNN
• CNN Batch Norm
• CNN LSTM CTC 2
• CNN LSTM CTC 1