Abstract:
Handwritten recognition systems take input from the paper documents scanner, images; touch
screen, and, electronic pen. After processing it offers the output as a electronic document suitable
for subsequent access and manipulation. There are a lot of handwritten documents. Which are
damaged due to dust, flooding, fire, and natural disaster like an earthquake. Nowadays there are a
lot of documents putting on shelves for many years and it takes more space it is bulky to manage
and edit a majority of data available in a handwritten form. This requires an approach to change
the documents into electronic form for easy of searching and retrieval as per users’ need.
In this research work, we have presented the development of offline handwritten character
recognition model for Awngi documents. We focus on the 26 base characters Awngi language.
The proposed character recognition model contains all of the essential steps that are obligatory for
developing an efficient recognition system. The designed model includes modules like
preprocessing, segmentation, feature extraction, and classification. In the preprocessing stage,
image resizes grayscale, noise reduction, morphological transformation, and binarization. In any
handwritten character recognition system is separating individual characters from the document.
Better character segmentation phase, has achieved using dual thresholding criterion to minimize
the character segmentation error. A CNN is used for feature extraction character classification
purposes.
Furthermore, In this paper, we prepared a new public image dataset for Awngi handwritten
characters a dataset containing a total of 30,115 out of the total dataset images,80%of dataset is
used for training, 20% dataset is used for testing. We collected data from injibara teacher’s
education, injibara baunk primary school students and Amhara mass media from Awngi media.
We are interested in the new success of end-to-end learning in pattern recognition we propose a
new trained end-to-end fashion model. The experimental results on handwritten Awngi character
recognition that our model achieves a training accuracy of 96.6% and 92.6% of the testing
accuracy of the proposed model.