Source-linked AI summary
Multi-Column Deep Neural Networks for Offline Handwritten Chinese Character Classification
Dan Cireşan, Jürgen Schmidhuber
TL;DR
Recognizing offline handwritten Chinese characters is difficult because the task has 3,755 classes and limited samples per class. The paper applies Multi-Column Deep Neural Networks to ICDAR competition data, correcting a preprocessing mismatch between training and submitted testing. The corrected systems approach human performance and improve substantially over single networks and prior artificial methods.
Problem
Offline Chinese handwriting recognition must handle 3,755 character classes, far more than digit-recognition tasks.
Method
The paper applies Multi-Column Deep Neural Networks to ICDAR 2011 and 2013 offline handwritten Chinese-character data.
Results
4.215% error is achieved by the best MCDNN versus 5.528% for the best DNN, while 4.21% approaches the measured human error rate of 3.87%.
Takeaways & Limitations
The MCDNN classifies 3,755 Chinese-character classes with almost human performance and is reported as ready for practical applications.
Takeaways & Limitations
Training and validation used preprocessing different from that in the submitted executable because Matlab and OpenCV routines produced different results and operation orders.
Abstract
from arXiv · showhide
Our Multi-Column Deep Neural Networks achieve best known recognition rates on Chinese characters from the ICDAR 2011 and 2013 offline handwriting competitions, approaching human performance.
IDSIA / USI-SUPSI
The paper is authored by Dan Cires¸an and J¨urgen Schmidhuber at IDSIA, affiliated with USI and SUPSI. The work received partial support from the Supervised Deep / Recurrent Nets SNF grant.
- IDSIA is affiliated with both USI and SUPSI.
- The work was partially supported by the Supervised Deep / Recurrent Nets SNF grant, Project Code 140399.
- The authors are Dan Cires¸an and J¨urgen Schmidhuber.
1 Introduction
Multi-Column Deep Neural Networks had previously achieved human-competitive digit recognition, and this work applies them to the substantially harder task of recognizing 3,755 offline handwritten Chinese character classes.
- Independent deep neural networks can be combined by output averaging into MCDNNs with error rates 20-40% below those of single networks.
- MCDNN achieved human-competitive performance on MNIST handwritten digit recognition in 2012.
- Chinese handwriting is harder than digit recognition because it involves 3,755 classes rather than 10.
- The paper applies MCDNN to ICDAR 2013 offline handwritten Chinese character data after correcting an image-preprocessing bug.
2 Details
The experiments use isolated offline Chinese-character images from ICDAR and HWDB 1.1, resized and contrast-maximized before classification by deep networks. A mismatch between training and submitted-test preprocessing affected the original competition results, which improved after correction.
- 2 Details: The datasets contain isolated offline Chinese characters, including 224419 competition test characters written by 60 persons.
- 2 Details: HWDB 1.1 has 3755 classes, with 897758 training characters from 240 persons and 223991 validation characters from 60 persons.
- 2.2 Preprocessing: Characters are uniformly rescaled into centered 48 × 48 pixel images after contrast maximization, using 40 × 40 pixels as the target character size.
- 2.3 Preprocessing glitch at ICDAR: The submitted executable used OpenCV preprocessing that differed from the Matlab routine used for training and validation, including scaling behavior and operation order.
- 2.3 Preprocessing glitch at ICDAR: 4.21% test error followed identical preprocessing for training and testing, versus 5.58% in the original competition result.
- 2.3 Preprocessing glitch at ICDAR: The corrected 2011 result was 5.78% error instead of 7.82%, a 2.04% lower error rate.
- 2.4 Network architecture: The networks use 11 layers, 3755 output neurons, and 100-450 maps per layer, with some trained on characters from all 300 associated persons.
3 Results
Multi-Column Deep Neural Networks substantially outperform single networks on the Chinese-character task, approach measured human performance, and retain useful GPU recognition speed.
- The best MCDNN reaches 4.215% error versus 5.528% for the best DNN, a 1.313% absolute and 23.75% relative reduction.
- MCDNNs consistently improve over single DNNs across the evaluated models.
- The best MCDNN obtains 4.21% error compared with the organizers’ measured human error rate of 3.87%.
- Its top-ten prediction error is 0.291%, a record reported as relevant to context-driven systems using linguistic models.
- The best MCDNN classifies 45 characters per second on a single NVIDIA GTX 580.
4 Conclusions and future work
MCDNNs classify 3,755 handwritten Chinese character classes with almost human performance and nearly one-fifth lower error than the best previous artificial method. GPU recognition is fast and scales linearly with the number of GPUs, while native-speaker error analysis could clarify remaining errors.
- 3,755 classes are classified with almost human performance, nearly one-fifth better than the best previous artificial method.
- GPU recognition speed is high and scales linearly with the number of GPUs.
- Native-speaker and native-writer error analysis could determine whether remaining errors reflect illegible characters or room for improvement.
- Additional context-driven linguistic models are expected to reduce errors further, and the method is described as ready for practical applications.