Executive Summary : | Multiclass object detection algorithms based on large monolithic models face limitations in frequent addition of newer categories and deletion or modification of previous categories. This is known as the incremental learning requirement in production systems, and the catastrophic forgetting problem, which occurs when a model loses prior performance upon retraining. One possible solution is to create multiple one class classifiers that operate on the input image simultaneously and perform post-processing to synthesize the final result. However, this line of addressing the problem has not been pursued. Optical character recognition (OCR) methods have evolved over several decades to address this purpose, covering printed and handwritten texts with multilingual features. The first two generations of OCR required rule-based segmentation followed by machine learning based classification based on proximity context features. The 3G OCR uses deep learning methods such as connectionist temporal classification, recurrent neural networks, and convolutional neural networks (CNN) based on vector sequence representation and mapping between pixels of input sub region and unicode sequence. Handwritten text recognition is another challenging sub-area of interest in the OCR community, as it requires accurate recognition of various factors such as cursive writing style, cluttering, font size variations, non-homogeneous orientation of sub-patterns, and background characteristics. Researchers are working on developing algorithms for creating lightweight OCCs for selective fusion, which involve fundamental mathematical formulation at the level of patterns, anti-patterns, sub-patterns, and inverse-patterns. This idea has been evaluated on OCR data sets and refinements are under research. |