Master thesis in Physics presentation. The title is "A Computational Model for Handwritten Text Detection and Segmentation in Historical Spanish Manuscripts".
Abstract
Historical manuscripts are worthy to be digitalized for analyzing its important history values, but it costs huge amount of time to recognize and transcribe manually. Applying the handwritten text recognition (HTR) technique into historical documents is necessary to improve the efficiency. For proceeding HTR accurately, it is essential to detect and segment the handwritten text region. This paper introduces a deep learning method to solve the problems of accurately identifying handwritten text areas. The model uses a deep convolutional neural network (CNN) architecture, ResNet, as encoder and specifies several different number of feature maps as decoder to predict the mask for handwritten text region detection and segmentation. The trained model was conducted on the dataset from the handwritten tabular census of Argentina in 1855. With the inference post-processing step, it can efficiently distinguish and segment the handwritten text area from the document that contains mixed-contents in both handwritten and printed.
Student's name: Xinlei Li
Supervisor: Dana Dannélls
Examiner: Mohsen Mirkhalaf