Table 12 Results of benchmark models along with proposed one
Sl. no | Benchmark model | Feature-types | Classifier | Dataset size | Test-sample count | Feature-vector size | Max. recg. acc. (%) for Hindi numerals |
|---|---|---|---|---|---|---|---|
1 | Khanduja et al. [3]a | Hybrid of structural and statistical features (intersection points, end points, loops, and pixel distributions) | MLP | 22,556 | 2000 | 462 | 95.5 |
2 | Trivedi et al. [9] | Image | CNN with genetic algorithm and L-BFGS method | 22,546 | 3762 | 256 | 96.54 |
3 | Kumar et al. [10]b | Image | Convolution autoencoder | 17,000 | 3400 | 1032 | 99.59 |
4 | Singh et al. [4] | Regional-weighted run length | MLP, NB, logistic, RF, SVM | 6000 | 2000 | 196 | 95.02 with SVM |
5 | Chaurasiya et al. [11]b | CNN-based features | SVM | 22,556 | 3759 | 1600 | 99.41 |
6 | Sarkhel et al. [12]b | CNN-based features | SVM | 3000 | 1000 | 4096, 2560, & 1792 | 99.5 |
7 | Proposed model | Fusion-based features as received from VGG-16Net, VGG-19Net, ResNet-50, and Inception-v3 | SVM | 20,000 | 5000 | 80 | 99.72 |