Table 6 MallaNet’s test accuracy on the DHCD compared to Previous work, with statistical significance assessed via McNemar’s test (\(p < 0.05\) denoting significant improvement) and approximate parameter counts derived from architecture descriptions or standard values where available.

From: MallaNet residual branch merge convolutional neural network with homogeneous filter capsules for Devanagari character recognition

Study

Model

Test accuracy (%)

Parameters

(approx. M)

Pal and Chaudhuri4

Gradient features + Quadratic classifier

94.80

N/A

Acharya et al.2

Deep CNN with dropout

98.47\(^\dagger\)

0.03

Aneja et al.8

Inception V3 (transfer learning)

99.00\(^\dagger\)

23.8

Mishra et al.1

ResNet-85 (fine-tuned pre-trained)

99.72

39

Masrat et al.3

Custom CNN

99.16\(^\dagger\)

N/A

Saini et al.9

Modified LeNet-5

99.21\(^\dagger\)

0.4

Mehta et al.10

Two-layer CNN

96.36 (36 classes)

N/A

Malla11

Hybrid quantum-classical CNN

99.80 (digits)

2.3

Proposed (MallaNet)

MallaNet

99.71

17

  1. Significant values are in [bold].
  2. \(^\dagger\)Statistically significant improvement by MallaNet (\(p < 0.05\), McNemar’s test).
  3. All studies use the DHCD with the standard 85/15 train-test split, though Mehta et al.10 and Malla11 evaluate on subsets; data augmentation and training regimes may vary.
Back to article page