Urdu is a complex language as it is an amalgam of many South Asian and East Asian languages; hence character recognition of Urdu is a huge and difficult task. We have systematically studied a recognition problem of handwritten Urdu digits. Our contribution is as follows:
- Firstly we visualize our dataset using PCA, t-SNE anda combination of PCA+t-SNE
- We apply 3 different architecture combinations namely:
- ResNet version 1 and version 2
- ResNet version 1 with Squeeze and Excitation block, ResNet version 2 with Squeeze and Exci-tation block
- ResNet version 1 with Convolutional Block Atten-tion Module, ResNet version 2 with ConvolutionalBlock Attention Module3)
Links to explore:
Result:
Attempt | Base Model | SE | CBAM |
---|---|---|---|
ResNetV1 | 97.76% | 98.3% | 98.68% |
ResNetV2 | 98.1% | 98.5% | 99.2% |
Confusion Matrix of ResNet V2 with CBAM