A Novel Approach Using Graph Convolutional Networks for Detecting Characters from Multi Language Scripts
Rahul Chiranjeevi. V, P. Ganesh Kumar, S Rubesh, M. G. · 2024
The projected market size for Optical Character Recognition (OCR) is in-creasing rapidly. The rise is propelled by the swift digitalization of company processes through the utilization of OCR technology, which aims to decrease labor expenses and conserve valuable work hours. While OCR is generally regarded as a resolved issue, one crucial aspect of it, recognizing hand writ-ten texts remain a tough subject. Different handwriting styles across various people and the low quality of text in comparison to printed text provide substantial challenges in the process of transforming it into machine-readable text. This research study presents different methods for offline handwriting character identification using a graph convolutional network. The process of character segmentation and normalization is expensive, particularly when dealing with big datasets, especially in the context of offline handwriting detection using a structural method. The primary objective of this research is to create a model that converts a handwritten character in-to a string graph representation. The goal of these models is to enhance recognition accuracy without depending on normalization techniques. The graph is composed of multiple edges that represent the inter-connected vertices. The vertices represent the curves that define the character's shape. The curve is derived from the analysis of the character's chain code, and its string characteristic is generated based on specific regulations. Proposed method achieved significant results compared to state of art methods in various benchmark datasets.