A convolution neural network with encoder-decoder applied to the study of Bengali letters classification
Sayed Mohsin Reza, Md Al Masum Bhuiyan, Nishat Tasnim · Big Data and Information Analytics · 2021
Handwritten grapheme recognition is popular research in computer vision and now widespread in the commercial industry due to its large number of applications in document analysis and recognition. Bengali grapheme classification is a complex task as it has 49 letters and 18 potential diacritics with almost 13,000 possible variations. Bengali is now the fifth most spoken native language and the seventh most spoken language by the total number of speakers in the world. Having set a bigger scope, this paper deals with the recognition of Bengali handwritten script letters. A class of deep Convolutional Neural Networks (CNNs) with encoder-decoder is used to classify handwritten letters. We use several serial non-linear layers as the encoders and a corresponding set of decoders that work as a pixel-wise classifier for letter recognition. The key idea is to encode images by convolution and decode them by deconvolution so that the max-pooling and up-sampling layers can correctly identify grapheme pixel-by-pixel. In this study, almost 200,840 grapheme images were analyzed, including root, vowels, and consonants. A large number of variations make additional complexity in recognition and may lead the model into over-fitting or under-fitting. We introduce regularization techniques to reduce the over-fitting in the fully connected layers. The results suggest that CNN with encoder-decoder can recognize complex grapheme characters with higher precision than traditional CNN. Experimental results show that the images' augmentation helps the model train better and improves its accuracy and loss.