BDNet: Bengali Handwritten Numeral Digit Recognition based on Densely connected Convolutional Neural Networks

Abstract

Bengali handwritten digit recognition can be done using different image classification techniques. But the images of handwritten digits are different from natural images as the orientation of a digit as well as similarity of features of different digits are important. On the other hand, deep convolutional neural networks are achieving huge success in computer vision problems, especially in image classification. This BDNet is a densely connected deep convolutional neural network model based on state-of-the-art algorithm DenseNet to classify Bengali handwritten numeral digits. The BDNet has end-to-end trained using ISI Bengali handwritten numeral dataset with 5-fold cross-validation. The BD-Net has achieved a test accuracy of 99.65%(baseline was 99.40%) on test data of ISI Bengali handwritten numerals. The trained model also gives 97.50% on own created dataset(which are not used during training). That is, this model gives a 41.66% error reduction compared to the previous state-of-the-art model. Codes, trained model and own dataset available at: https://github.com/Sufianlab/BDNet.

Two Decades of Bengali Handwritten Digit Recognition: A Survey

“…Recently, [58] proposed a densely connected deep CNN architecture, inspired by DenseNet architecture [145], called 'BDNet', which achieved 99.78% accuracy on the ISI-HBN dataset. The input is passed through a series of 'Dense' and 'Transition' blocks consisting of several 2D convolutions (Conv2D), ReLU, and Batch Normalization (BN) layers.…”

“…As the size, variation, and background of the samples in the dataset got diversified, the authors have used complex models with more layers. For example, since the NumtaDB dataset contains different challenging samples, most of the works have utilized CNN models with more than six layers to train their models to achieve acceptable performance on this dataset [42], [54], [56]- [58], [62], [63], [166].…”

“…This process was repeated a couple of times. Reference [58] used an LR of 0.009 till 80 epochs and then reduced it by a factor of 0.15 until the model converged. Reference [62] used two different LRs (0.001 and 0.0001) to train the ensembled CNN networks.…”

A new feature extraction approach for script invariant handwritten numeral recognition

“…Recently, convolutional neural networks (CNNs) have emerged as a promising deep learning based approach for the recognition of handwritten numerals (Keserwani et al, 2019;Rabby et al, 2019;Sufian et al, 2020). The work described by Ashiquzzaman and Tushar (2017) proposed a method based on a CNN which achieved an accuracy of 97.4% when tested on the CMATERdb3.3.1 (CMATER Dataset, 2018) Arabic numeral database.…”

“…The work described by Ashiquzzaman and Tushar (2017) proposed a method based on a CNN which achieved an accuracy of 97.4% when tested on the CMATERdb3.3.1 (CMATER Dataset, 2018) Arabic numeral database. For Bangla handwritten numeral recognition, Sufian et al (2020) introduced densely connected CNNs named BDNet, which was tested on ISI Bengali handwritten numerals and yielded an accuracy of 99.78%. Reddy et al (2018) designed an efficient CNN based on Root Mean Square Propagation (RMSprop) optimization for Devanagari numeral recognition.…”