Department of Electronics and Communication Engineering

Permanent URI for this collection

http://nits.ndl.gov.in/handle/123456789/91

Browse

Browsing Department of Electronics and Communication Engineering by Author "Saidulu, Naragoni"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Handwritten Character Recognition using Deep Convolutional Neural Networks
    (Department of Electronics and Communication Engineering, NITS, 2022) Saidulu, Naragoni
    Human perception takes the help of size information to recognize the upper- and lower-case alphabets. Recognizing handwritten characters is difficult mainly because of the variation in the character pattern, handwriting style, preferred size, and larger dictionary. To handle this, researchers have opted for designing and developing customized networks like CapsuleNet, TextCaps, etc. Also, the performance of these networks on case-sensitive characters is mini mal, because of the inability to distinguish characters that are differentiable only by size and not by their pattern. The results of the handwritten character recognition (HCR) based open source state-of-art (SOTA) deep convolutional neural networks reveal that the performance is improved on testing NIST SD19 (image:128x128) compared to EMNIST (image: 28x28). On equivalent, to human perception, the developed SOTA-DRBAM-Net (state-of-art deep residual bottleneck attention convolutional neural network) proved that channel and spatial attention at the bottleneck of each residual block helps to improve the performance further. Also, SOTA DRBAM-Net can (i) have a recognition performance more than the existing state-of-the-art network; (ii) resolve and handle the complexity of case-sensitiveness in handwritten characters provided peripheral information is being preserved. The performance of SOTA-DRBAM-Net on (i) EMNIST (By-class: accuracy (AR)- 88.05%, By-merge: AR– 91.2%, Balanced: AR 89.6%, Letters: 94.8%, Digits: 99.6%) (ii) NIST SD19 (By-class: AR- 91.70%, By-merge: AR– 92.86%, Balanced: AR-90.89%, Letters: 90.88%, Digits: 99.69