Introduction
Online Handwritten Chinese Character Recognition (HCCR) is a significant challenge due to the vast character set—standard Chinese involves over 3,700 frequently used characters. Unlike static image recognition, "online" recognition tracks the actual trajectory of the pen over time. While deep learning has pushed recognition accuracy to near-human levels, the resulting models are often massive, requiring between 70 MB and 135 MB of storage. This creates a significant "memory tax" for mobile and embedded devices. This paper presents a method to compress these massive Convolutional Neural Networks (CNNs) down to just 0.57 MB—a 210x reduction compared to traditional methods.
Methodology
The researchers achieved extreme compression through three primary strategies:
1. Path Signatures
To represent handwriting effectively, the system uses "path signatures." These are mathematical iterated integrals that capture the displacement and curvature of pen strokes over time, providing a rich, multi-dimensional input for the CNN.
2. Global Average Pooling (GAP)
In traditional CNNs, the final layers are "Fully Connected" (FC), which contain the vast majority of the model's parameters. The authors replaced these bulky FC layers with GAP layers. Instead of millions of weights, GAP simply takes the average of each feature map, drastically reducing the parameter count. While GAP initially shows slightly lower accuracy in dense models, it proved much more resilient to aggressive pruning than FC layers.
3. DropWeight Pruning
Standard pruning usually involves cutting all neural connections below a certain weight threshold. However, setting a high threshold too early can prevent model recovery. The authors proposed DropWeight, a technique where the pruning threshold is increased gradually over many iterations. This allows the network to "re-learn" and compensate for lost connections incrementally.
Key Results
The models were evaluated on the ICDAR-2013 online HCCR competition dataset (3,755 classes). The results demonstrate a masterclass in trade-off management between size and performance.
As shown in the table, the best compact model (HCCR-Inc-GAP-Pruned) shrunk to a mere 0.57 MB while maintaining an accuracy of 96.88%. Compared to the previous state-of-the-art "DropDistortion" model (19.03 MB), the proposed model is roughly 33 times smaller with only a 0.91% dip in performance. Compared to traditional non-CNN methods, it is over 200 times smaller.
Conclusion
The research proves that standard models contain immense redundancy. By combining architecture-agnostic pruning with Global Average Pooling, high-performance handwriting recognition becomes viable for even the most resource-constrained mobile devices. Future work will focus on extending these compression techniques to Recurrent Neural Networks (RNNs) for continuous text recognition.
Summary updated