End-To-End Deep Learning Based Tamil Handwritten Document Recognition
and Classification Model
Abstract
Handwritten recognition (HR) remains a challenging process in various
real-world applications. Tamil handwritten text recognition involves the
recognition of text in scanned images. Recognition of handwritten Tamil
characters is a tedious process because of the differences in sizes,
style and orientation angle. Prior studies concentrated on
character-level segmentation and each character was subsequently
classified. Segmentation is then used, first at the word level and
subsequently at the line level. The recently developed machine learning
(ML) and deep learning (DL) approaches can be utilized for Tamil HCR.
With this motivation, this paper presents an end-to-end deep
learning-enabled Tamil handwritten document recognition (ETEDL-THDR)
model. The ETEDL-THDR paragraph text recognition can be accomplished by
the use of two modules such as line segmentation and line recognition.
Initially, the ETEDL-THDR model enables the improvement of the quality
of the input images by the use of the median filtering (MF) technique.
To create meaningful regions, further line and character segmentation
activities are performed. Additionally, a deep convolutional neural
network-based MobileNet approach was applied to derive feature vectors.
At last, the water strider optimization (WSO) algorithm with a
bidirectional gated recurrent unit (BiGRU) model is applied to recognize
Tamil characters. An extensive experimental analysis of the ETEDL-THDR
model is carried out and the results showed that the ETEDL-THDR model
performed better than more recent methodologies with maximum accuracy of
98.48%, precision of 98.38%, a sensitivity of 97.98%, specificity of
98.27%, and F-measure of 98.35%.