Tamil OCR Conversion from Digital Writing Pad Recognition Accuracy Improves through Modified Deep Learning Architectures

Table 1

Comparison of different methods and languages.

References	Model	Language	Dataset	Type	Dataset size	Writing devices
[16]	CNN	Arabic	CMATERDB	Digits	3,000	Pen and Paper
[17]	CNN BiLSTM	Arabic	KHATT	Word	325 paragraph	Pen and Paper
[18]	CNN-TL	Bangla and Devanagiri	CMATERDB ICDAR Ph.D. Indic_11	Word	361 handwritten documents	Pen and Paper
[19]	GAN	Chinese	HIT-MW and ICDAR	Word	1,003 handwritten images	Pen and Paper
[20]	CNN	Telugu	Own dataset	Character	16 characters and 21 guninthalu, total of 275,520	Pen and Paper Touchpad device
Proposed	RTSBA	Tamil	DWP-H and HP Lab datasets	Word	251 writers total of 85,800 images	Tablet

CNN, convolutional neural network; BiLSTM, bidirectional long short-term memory; RTSBA, ResNet two-stage bottleneck architecture; DWP-H, digital writing pad-handwritten; GAN, generative adversarial network; CMATERDB, center for microprocessor applications for training education and research data base; KHATT, KFUPM Handwritten Arabic TexT; ICDAR, international conference on document analysis and recognition; HIT-MW, harbin institute of technology-multiple writers.