[Paper Review] DropSample: A New Training Method to Enhance Deep Convolutional Neural Networks for Large-Scale Unconstrained Handwritten Chinese Character Recognition

Weixin Yang, Lianwen Jin+3 more|arXiv (Cornell University)|May 20, 2015

Handwritten Text Recognition Techniques57 references4 citations

TL;DR

This paper introduces DropSample, a novel training method for deep convolutional neural networks (DCNNs) that dynamically reweights training samples based on classification confidence—prioritizing low-confidence samples to improve learning efficiency. Applied to large-scale unconstrained handwritten Chinese character recognition, DropSample with domain-specific knowledge layers achieved state-of-the-art accuracies of 97.33%, 97.06%, and 97.51% on CASIA-OLHDWB 1.0, CASIA-OLHWDB 1.1, and ICDAR 2013 datasets, respectively.

ABSTRACT

Inspired by the theory of Leitners learning box from the field of psychology, we propose DropSample, a new method for training deep convolutional neural networks (DCNNs), and apply it to large-scale online handwritten Chinese character recognition (HCCR). According to the principle of DropSample, each training sample is associated with a quota function that is dynamically adjusted on the basis of the classification confidence given by the DCNN softmax output. After a learning iteration, samples with low confidence will have a higher probability of being selected as training data in the next iteration; in contrast, well-trained and well-recognized samples with very high confidence will have a lower probability of being involved in the next training iteration and can be gradually eliminated. As a result, the learning process becomes more efficient as it progresses. Furthermore, we investigate the use of domain-specific knowledge to enhance the performance of DCNN by adding a domain knowledge layer before the traditional CNN. By adopting DropSample together with different types of domain-specific knowledge, the accuracy of HCCR can be improved efficiently. Experiments on the CASIA-OLHDWB 1.0, CASIA-OLHWDB 1.1, and ICDAR 2013 online HCCR competition datasets yield outstanding recognition rates of 97.33%, 97.06%, and 97.51% respectively, all of which are significantly better than the previous best results reported in the literature.

Motivation & Objective

• To improve the efficiency and accuracy of deep convolutional neural networks in large-scale unconstrained handwritten Chinese character recognition (HCCR).
• To address the challenge of inefficient learning in DCNNs due to redundant training on already well-recognized samples.
• To develop a dynamic sampling strategy that adaptively focuses training on difficult, low-confidence samples.
• To integrate domain-specific knowledge into the CNN architecture to further boost recognition performance.

Proposed method

• DropSample introduces a quota function per training sample that is dynamically adjusted based on the DCNN's softmax confidence output after each learning iteration.
• Samples with low classification confidence are assigned higher probabilities of selection in subsequent training iterations, increasing their learning impact.
• Well-classified, high-confidence samples are gradually deprioritized and removed from training to reduce redundancy and improve efficiency.
• A domain knowledge layer is inserted before the traditional CNN to inject task-specific invariants and structural priors into the network.
• The method combines dynamic sampling with architectural inductive bias to enhance feature learning and generalization.
• The training process iteratively refines model capacity by focusing on hard examples, mimicking the Leitner learning box principle from cognitive psychology.

Experimental results

Research questions

• RQ1 Can dynamic reweighting of training samples based on model confidence improve the convergence and accuracy of DCNNs in HCCR?
• RQ2 How does prioritizing low-confidence samples affect the learning efficiency and final recognition performance?
• RQ3 To what extent can domain-specific knowledge enhance DCNN performance in unconstrained handwritten Chinese character recognition?
• RQ4 Does the combination of dynamic sampling and architectural inductive bias yield superior results compared to standard training protocols?

Key findings

• DropSample achieved a recognition accuracy of 97.33% on the CASIA-OLHDWB 1.0 dataset, surpassing previous state-of-the-art results.
• On the CASIA-OLHWDB 1.1 dataset, the method reached 97.06% accuracy, demonstrating robustness across different data distributions.
• The ICDAR 2013 online HCCR benchmark was solved with a 97.51% recognition rate, setting a new SOTA performance.
• The integration of domain-specific knowledge with DropSample significantly improved model generalization and convergence speed.
• The dynamic sampling mechanism reduced training redundancy by deprioritizing high-confidence samples, leading to faster and more efficient learning.
• The method's effectiveness was validated across three diverse, large-scale HCCR benchmarks, confirming its generalizability.