Source-linked AI summary
High Performance Offline Handwritten Chinese Character Recognition Using GoogLeNet and Directional Feature Maps
Zhuoyao Zhong, Lianwen Jin, Zecheng Xie
TL;DR
Offline HCCR remains challenging because prior CNNs were not sufficiently deep or slim and often omitted useful domain-specific features. The paper introduces a streamlined GoogLeNet with directional feature maps, achieving 96.35% single-model and 96.74% ensemble recognition accuracy on the ICDAR 2013 dataset.
Problem
Existing HCCR CNNs were neither deep enough nor slim enough, while traditional directional features were not fully incorporated into end-to-end CNNs.
Method
The paper adapts GoogLeNet into a streamlined HCCR model and adds Gabor, gradient, and HoG directional feature maps to the input.
Results
96.35% single-model and 96.74% ensemble recognition accuracy were achieved on the ICDAR 2013 offline HCCR competition dataset.
Takeaways & Limitations
A deeper, slimmer GoogLeNet architecture combined with directional feature maps improves offline HCCR recognition performance.
Abstract
from arXiv · showhide
Just like its great success in solving many computer vision problems, the convolutional neural networks (CNN) provided new end-to-end approach to handwritten Chinese character recognition (HCCR) with very promising results in recent years. However, previous CNNs so far proposed for HCCR were neither deep enough nor slim enough. We show in this paper that, a deeper architecture can benefit HCCR a lot to achieve higher performance, meanwhile can be designed with less parameters. We also show that the traditional feature extraction methods, such as Gabor or gradient feature maps, are still useful for enhancing the performance of CNN. We design a streamlined version of GoogLeNet [13], which was original proposed for image classification in recent years with very deep architecture, for HCCR (denoted as HCCR-GoogLeNet). The HCCR-GoogLeNet we used is 19 layers deep but involves with only 7.26 million parameters. Experiments were conducted using the ICDAR 2013 offline HCCR competition dataset. It has been shown that with the proper incorporation with traditional directional feature maps, the proposed single and ensemble HCCR-GoogLeNet models achieve new state of the art recognition accuracy of 96.35% and 96.74%, respectively, outperforming previous best result with significant gap.
I. INTRODUCTION
Offline HCCR remains challenging because of its large vocabulary, varied handwriting styles, and confusable characters. The paper motivates deeper, slimmer CNNs that also incorporate useful domain-specific feature information.
- Motivation: HCCR remains difficult because Chinese character vocabularies are large, handwriting styles vary widely, and many characters are confusable.The cited standards contain thousands to tens of thousands of classes.
- Motivation: Traditional MQDF- and DLQDF-based methods appear to have reached a performance bottleneck.The passage reports no significant recent progress for these traditional approaches.
- Prior progress: CNNs have brought major advances to HCCR and narrowed the gap between recognition systems and human performance.Prior CNN work established a successful end-to-end approach for handwritten Chinese character recognition.
- Research gap: Many existing CNN-based HCCR models process characters as image patterns without explicitly using domain-specific feature extraction.The paper identifies this omission as a limitation of end-to-end black-box processing.
- Contribution: The paper proposes a streamlined GoogLeNet-based CNN and incorporates Gabor, gradient, and HoG directional feature maps.The feature maps are added alongside the original image at the CNN input.
III. DESIGN OF TWO CNNS FOR HCCR
The paper designs two CNN architectures for offline HCCR: an AlexNet-based model and a deeper GoogLeNet-based model using Inception modules.
- HCCR-AlexNet: The HCCR-AlexNet follows AlexNet and contains eight weighted layers, including convolutional, pooling, and fully connected layers.Its detailed architecture is presented in Figure 1.
- HCCR-GoogLeNet: HCCR-GoogLeNet follows GoogLeNet’s deep architecture and local Inception module.The Inception module seeks an optimal local construction that can be repeated spatially.
- HCCR-GoogLeNet: GoogLeNet is described as 22 layers deep when counting parameterized layers, or 27 layers when pooling layers are included.These depths characterize the original GoogLeNet architecture motivating the proposed model.
IV. EMBEDDING OF DIRECTIONAL FEATURE MAPS TO HCCR-GOOGLENET
The method augments the original character image with directional feature maps as prior knowledge for HCCR-GoogLeNet. It uses Gabor, gradient, and HoG representations to provide complementary feature inputs.
- Feature-map integration: Directional feature maps are added to the input layer alongside the original image to enhance HCCR-GoogLeNet.The paper frames these maps as domain-specific prior knowledge that CNNs may not learn directly.
- Gabor features: Gabor transformation produces orientation-specific feature maps by filtering the input image with multi-orientation Gabor filters.The transformation is described as a convolution of the input image with an oriented Gabor filter.
- Gabor features: The Gabor configuration uses eight orientations, and the resulting eight maps are concatenated with the original image into a 9-channel CNN input.The orientations range from 0° to 157.5° in 22.5° increments.
- Gradient and HoG features: Gradient maps are computed with Sobel operators on the x- and y-axes and decomposed into eight directional maps.These maps are inserted into the input layer with the primary image.
- Gradient and HoG features: HoG features are also used as an additional directional representation for HCCR.The paper illustrates Gabor, gradient, and HoG maps for the handwritten character “积”.
A. Experimental Data
The experiments train on two CASIA handwritten-character databases and test on the ICDAR 2013 CompetitionDB dataset.
- Datasets: Training uses CASIA-HWDB1.0 and CASIA-HWDB1.1, both collected by the Institute of Automation of the Chinese Academy of Sciences.The databases provide Chinese-character classes contributed by multiple writers.
- Datasets: Testing uses CompetitionDB from the 2013 ICDAR Chinese handwriting recognition competition.The test dataset is also associated with the Institute of Automation of the Chinese Academy of Sciences.
B. Pre-processing and experimental settings
The experiments normalize offline character images to model-specific sizes and center scaled characters within larger masks to preserve margin information.
- Images are normalized to 108x108 for HCCR-AlexNet and 112x112 for HCCR-GoogLeNet.Grey values are reversed before resizing for faster computation.
- Scaled characters are centered in 114x114 and 120x120 masks for HCCR-AlexNet and HCCR-GoogLeNet, respectively.The larger masks maintain margin information.
- Training data are shuffled before preprocessing and experimentation.
C. Comparison of HCCR-GoogLeNet against HCCR-AlexNet
HCCR-GoogLeNet substantially outperforms HCCR-AlexNet on the 3755-class CompetitionDB test set, supporting the benefit of its deeper architecture.
- 96.26% recognition accuracy was achieved by HCCR-GoogLeNet, compared with 95.49% for HCCR-AlexNet.The comparison uses the testing 3755 classes of CompetitionDB.
- The deeper HCCR-GoogLeNet architecture is reported to extract more natural and abstract character representations and improve accuracy.
D. Results of HCCR-GoogLeNet plus Gabor for HCCR
Directional feature maps improve HCCR-GoogLeNet performance, with Gabor features producing the strongest single-feature result and ensembles achieving higher accuracy.
- 96.35% recognition accuracy was achieved by Gabor+HCCR-GoogLeNet, exceeding the gradient, HoG, and plain HCCR-GoogLeNet variants.The experiment embeds directional feature maps into the input layer.
- 95.08% recognition accuracy was obtained using only 8-directional Gabor feature maps without the original bitmap.This result demonstrates character representation through Gabor feature maps alone.
- 96.64% and 96.74% accuracy were obtained by combining four and ten pretrained HCCR-GoogLeNet models, respectively.
E. Comparison of different methods on the CompetitionDB
On the ICDAR 2013 offline HCCR competition dataset, HCCR-GoogLeNet outperforms traditional, competition-winning, and prior CNN-based methods while using a deeper but slimmer model.
- 55.22% relative error-rate reduction was achieved against the best traditional MQDF-based DLQDF result.
- 37.67% relative error-rate reduction was achieved against CNN-Fujitsu, the ICDAR 2013 competition winner.
- 17.26% relative error-rate reduction was achieved against the best state-of-the-art CNN result, ATR-CNN Voting.
- HCCR-GoogLeNet uses 19 layers and 7.26 million parameters, compared with ATR-CNN's 10 layers and approximately 12.91 million parameters.The reported storage sizes are 27.68MB for HCCR-GoogLeNet and 51.64MB for ATR-CNN.
VI. CONCLUSION
The paper presents HCCR-GoogLeNet, a deep and slim network that incorporates directional feature extraction for handwritten Chinese character recognition. On the ICDAR 2013 offline HCCR dataset, it achieves a best testing error rate of 3.26%, reported as a new state-of-the-art record.
- HCCR-GoogLeNet uses four Inception modules to construct an efficient deep network for handwritten Chinese character recognition.The modules support repeated spatial construction of local network structures.
- The model has 19 layers while remaining slim, addressing the paper’s emphasis on deep yet efficient HCCR architectures.
- Directional Gabor, HoG, and gradient features are incorporated as domain knowledge to enhance HCCR-GoogLeNet’s performance.
- 3.26% is the best testing error rate achieved on the ICDAR 2013 offline HCCR competition dataset.The authors report this as a new state-of-the-art record to their best knowledge.
- The best single HCCR-GoogLeNet surpasses previous best single and ensemble CNN models in both accuracy and storage performance.