A study on the applicability of Long Short-Term Memory networks to industrial OCR

This thesis summarises the research-oriented study of applicability of Long Short-Term Memory Recurrent Neural Networks (LSTMs) to industrial Optical Character Recognition (OCR) problems. Traditionally solved through Convolutional Neural Network-based approaches (CNNs), the reported work aims to detect the OCR aspects that could be improved by exploiting recurrent patterns among pixel intensities, and speed up the overall character detection process. Accuracy, speed and complexity act as the main key performance indicators. After studying the core Deep Learning foundations, the best training technique to fit this problem first, and the best parametrisation next, have been selected. A set of tests eventually validated the preciseness of this solution. The final results highlight how difficult is to perform better than CNNs for what OCR tasks are concerned. Nonetheless, with favourable background conditions, the proposed LSTM-based approach is capable of reaching a comparable accuracy rate in (potentially) less time.