Handwritten Gujarati Numerals Classification Based on Deep Convolution Neural Networks Using Transfer Learning Scenarios

In recent years, handwritten numeral classification has achieved remarkable attention in the field of computer vision. Handwritten numbers are difficult to recognize due to the different writing styles of individuals. In a multilingual country like India, negligible research attempts have been carried out for handwritten Gujarati numerals recognition using deep learning techniques compared to the other regional scripts. The Gujarati digit dataset is not available publicly and deep learning requires a large amount of labeled data for the training of the models. If the number of annotated data is not sufficient enough to train Convolutional Neural Networks (CNN) from the scratch, transfer learning can be applied. However, the issue arises by using transfer learning is that how deep to fine-tune the pre-trained convolutional neural network while training the target model. In this paper, we addressed these problems using three deep transfer learning scenarios to classify handwritten Gujarati numerals from the images of zero to nine. We presented transfer learning scenarios using ten pre-trained CNN architectures including LeNet, VGG16, InceptionV3, ResNet50, Xception, ResNet101, MobileNet, MobileNetV2, DenseNet169 and EfficientNetV2S to find the best performing model by freezing and fine-tuning the weight parameters. We implemented the pre-trained models using a self-created handwritten Gujarati digit dataset with 8000 images of zero to nine digits with data augmentation. Exhaustive experiments are performed using various performance evaluation matrices. EfficientNetV2S model showed promising results among all the models including three transfer learning scenarios and achieved 98.39% training accuracy, 97.92% testing accuracy, 97.69% f1-score, and 97.15% AUC. Our handwritten Gujarati digit dataset is available on https://github.com/Parth-Goel/gujarati-handwritten-digit-dataset/.

Automated summary

This study addresses the problem of handwritten Gujarati numeral recognition in India. It uses deep transfer learning to find the best performing model by freezing and fine-tuning the weight parameters of ten pre-trained CNN architectures.