Development of Convolutional Recurrent Neural Network (CRNN) deep learning model for Indonesian optical character recognition back end web service
This thesis researches and experiments on developing a web service for Indonesian Optical Character Recognition (OCR) by adjusting some of the affecting hyper-parameters and structuring layer. The research focuses on how each parameter chosen have optimum effects on the prediction result of the printed text. This topic is relevant since the current model available are trained based on English documents which leads the predictions biased towards the English language. This research uses Connectionist Temporal Classification (CTC) in order to determine how each label corresponds to the text in the image instead of segmenting the characters one by one. Thus, the usage Convolutional Recurrent Neural Network (CRNN). The first experiment uses augmented generated text in order to eliminate the need of data factor. The second experiment uses real data set in order to verify the results. This experiment finds that the hyper-parameter values that produces the preferred accuracy are GRU for sequencing layer, 60 for epoch value, 32 for batch size, and 0.01 as learning rate. The final results achieved for Surat Kuasa, Kartu Tanda Pengenal, Kartu Keluarga are 98%, 95%, 97% respectively while the character accuracy for those documents are 78% ,67%, 62% respectively.
|B03000||(Rack Thesis)||Available but not for loan - Missing|
No other version available