For more reference, and related explaination please refer to medium article: American Sign Langugage recognition using CNN Different CNN structures for classifying American Hand Signs along with their training and testing accuracies:
Model | Number of Convolution layers | Augmentation | Batch Normalisation | Dropout | Training Accuracy | Test Accuracy |
---|---|---|---|---|---|---|
model1 | 1 | yes | yes | no | 97.52 | 97.71 |
model2 | 2 | no | no | no | 100 | 91.06 |
model3 | 2 | no | yes | no | 100 | 95.60 |
model4 | 2 | yes | yes | no | 94.58 | 98.41 |
model5 | 2 | yes | yes | no | 99.32 | 99.71 |
model6 | 2 | yes | yes | no | 98.70 | 99.51 |
model7 | 2 | yes | yes | 0.4 | 94.09 | 98.42 |
model8 | 2 | yes | yes | 0.4,0.4 | 84.27 | 91.98 |
model9 | 3 | yes | yes | no | 99.23 | 99.33 |
model10 | 3 | yes | yes | no | 91.90 | 98.1 |
model10 | 3 | yes | yes | 0.4 | 91.90 | 98.1 |
model11 | 3 | yes | yes | 0.2 | 92.32 | 97.5 |
model12 | 4 | no | yes | no | 100 | 96.10 |
model13 | 4 | yes | yes | no | 99.28 | 99.83 |
model14 with dynamic learning rate | 4 | yes | yes | no | 99.71 | 100 |