Comparative Study of Deep Learning models to Recognize Palm Leaf Malayalam Characters

Palm-leaf manuscripts are historical records written mainly in many local languages in many South Asian countries. These texts have suffered damage over many years and centuries. Many organizations have scanned and digitally imaged palm-leaf manuscripts to preserve these potentially priceless records. Machine readable text conversion should be done on these images for the easy retrieval and processing. A comparative study of deep learning models for recognizing palm-leaf Malayalam characters (PLMC) was conducted in this work. Characters were segmented directly from palm leaves using the contour method. Various CNN architectures were employed to recognize these segmented characters, namely, VGG16, LeNet, MobileNet, Resnet152v2, VGG19, and DenseNet121. Additionally, a custom CNN architecture was developed and experimented with. It was observed that the custom CNN architecture provides the best recognition accuracy at 96%.