HUTNet: An Efficient Convolutional Neural Network for Handwritten Uchen Tibetan Character Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Handwritten Uchen Tibetan character recognition is crucial for digital data acquisition but faces challenges like touching letters and similar character forms.
• Deeper neural networks are needed for accuracy, but this increases model size and latency, creating a trade-off.

Purpose of the Study:

• To design an efficient and lightweight neural network model for handwritten Uchen Tibetan character recognition.
• To balance accuracy and latency for practical applications on low-power devices.

Main Methods:

• Introduced HUTNet, a novel network architecture considering the relationship between floating-point operations per second (FLOPs) and Memory Access Cost.
• Applied pruning and knowledge distillation techniques to further optimize the model.
• Evaluated performance on the Handwritten Uchen Tibetan Data set by Wang (HUTDW) and Chinese Handwriting Database 1.1 (HWDB1.1).

Main Results:

• HUTNet achieved 96.86% accuracy on HUTDW with only 10% of ResNet-18's parameters.
• The compressed HUTNet model reached 96.83% accuracy on HUTDW, with reduced FLOPs and parameters.
• On HWDB1.1, HUTNet surpassed ResNet-18 and ResNet-34 with 97.24% accuracy.

Conclusions:

• HUTNet offers a superior balance between accuracy and computational resources compared to existing models.
• The model effectively addresses the bottleneck for handwritten Uchen Tibetan recognition on resource-constrained devices.
• HUTNet demonstrates strong performance and efficiency for handwritten character recognition tasks.