Historical digit recognition using CNN: a study with English handwritten digits

Handwriting-based technologies have progressed significantly over the years. Scientists have worked beyond just recognizing pieces of plain text from paper. With archaeological advancements, discovery of ancient documents has not been as scarce as it was in the past. However, such documents are discovered in perfect condition once in a blue-moon. They are often subjected to degradation due to the perils of time and understanding them requires skilled manpower which is not easy to find. Here, an automated system is proposed for this task. The proposed system uses a new CNN-based framework to analyze the macro and micro features of an image to recognize the text. Experiments are performed with over 250K historical images of numerals from a publicly available dataset and a highest accuracy of 99.68% is obtained. The data is further subjected to different type of noises and distortions. A novel technique is used to introduce synthesized degradation in the documents. The system performs steadily in all these experimental scenarios. Comparative analysis indicates that our results are higher than the reported works and other standard techniques.