Handwritten Bangla Text Recognition Using Deep Learning Approach

For transition from analog to digital systems, most of the documents are also turning in to digital system. Digital systems are easier to sort, merge, etc and have many automation support like sorting, merging, searching, context extraction, et cetera (etc.). Yet there is a large number of handwritten documents generated daily. Furthermore, a large amount daily information such as: bank cheques,  postal cards, government fillup forms like pass-port verification, license verification, etc. are highly depended on handwritten form. This documents are needed to handle manually and is very labour intensive. For switching to digital systems, there is also becoming a need to convert this handwritten documents to digitalized form by hand which is conventional manual input of data. Handwritten recognition has become very popular in this context. Bangla being one of the popular languages in the world are being used and written and a humongous amount of handwritten text documents are in need of process. However, bangla handwritten frameworks are very few. Available methods are also prone to low performance metrics as detecting Bangla handwritten texts are very complex in nature. Major challenges in this pattern recognition problem are: 1) A large amount of classes available in Bangla containing 50 basic characters, 10 digits, 15 diacrtics, and more than 250 compound characters while English and German languages has a few amount of classes to distinguish. 2) Bangla characters are cursive in nature and becomes a complex pattern for distinguishable features. 3) Many Bangla characters are analogous to each other making it very hard to differentiate between them. 4) The strange formulation of compound characters based on basic characters even makes this structure more complex in nature. In this study, we are proposing a Bangla handwritten recognition system based on Visual Geometry Group 16(VGG16) model. We presented a novel improved VGG16 model by implementing attention mechanism and learning rate controlling, regularization techniques. For attention mechanism we have implemented SE block while step decay is implemented as a learning rate controller. For regularization purposed batch normalization has been implemented. Instead of relying on pre-trained weights from unrelated datasets our proposed architecture has been implemented using three different datasets Banglalekha Isolated (BLI), Ekush, NumtaDB ensuring domain specific training and the framework maintained high accuracy in all of them. This consistent performance indicates our models versatility over different datasets and non-dependency on any particular datasets. Additionally, our model performed high accuracy metrics in other patterns such as compound characters, digits, and diacritics which illustrates our models robustness over other type of pattern recognition on handwritten tasks. The model achieved most accuracy with NumtaDb dataset of 98.79%. Our model showed stable accuracy on training and validation trend analysis on all the  datasets. 

• Supervision