TL;DR
This paper introduces Maximum Entropy Regularization (MER) to Chinese text recognition, a field plagued by massive character sets and extreme class imbalance. By simply adding a negative entropy term to the standard cross-entropy loss, the authors prevent models from becoming "overconfident," leading to better generalization, cleaner attention maps, and significantly higher robustness against noisy data.
What Problem Does This Paper Address?
Recognizing Chinese text is a "fine-grained" classification nightmare. Unlike Latin scripts with a few dozen characters, Chinese involves thousands of unique characters, many of which look nearly identical (high inter-class similarity) while appearing in vastly different handwritten or artistic styles (high intra-class variance).
Because some characters appear much more frequently than others, deep learning models tend to overfit. They become "too certain" about their predictions, collapsing the output probability distribution into a one-hot vector. This "overconfidence" causes the model to ignore subtle features that distinguish similar characters and makes it brittle when it encounters noisy backgrounds or mislabeled training data.
Methods
The authors propose a deceptively simple fix: Maximum Entropy Regularization (MER).
Standard training uses Cross-Entropy (CE) loss, which essentially "pushes" the model to assign a probability of 1.0 to the correct class. The authors add a second term to the loss function: $$L_{REG} = L_{CE} + \lambda L_{MER}$$ Where $L_{MER}$ is the negative entropy of the predicted distribution.
The Intuition: Think of these as two competing forces. Cross-entropy is a force trying to make the model certain (one-hot distribution). MER is a force trying to make the model uncertain (uniform distribution). When balanced by the hyperparameter $\lambda$, the model converges to a point where it is confident enough to be correct, but "leaves enough room" in the probability space for similar-looking characters.
Crucially, this method requires zero additional parameters and no changes to the model architecture (they used SE-ResNet50 and an attention-based Encoder-Decoder).
Datasets
The authors evaluated their approach on three challenging benchmarks:
- CTW (Chinese Text in the Wild): A massive dataset of street view images containing over 1 million characters across 3,850 unique classes.
- ReCTS (Reading Chinese Text on Signboard): Focused on signboard text, featuring 4,435 unique characters and over 100,000 text lines.
- CUB-200-2011: A classic fine-grained image classification dataset (birds). While not OCR, it was used to prove that MER works for any task involving high similarity between classes.
Key Insights and Ideas
- The "Force" Analogy: The authors provide a beautiful theoretical analysis showing that the gradient of the regularized loss acts like a physical balance. This allows them to mathematically predict the "convergence probability"—exactly how confident the model will be for a given $\lambda$.
- Better Attention, Less Noise: Visualization via activation maps (CAM) showed that models trained with MER focus on more compact, discriminative regions of a character. Models without MER often "over-attend" to noisy backgrounds just to find enough clues to reach that 100% confidence mark.
- MER vs. Label Smoothing: While Label Smoothing (LS) also prevents overconfidence, it does so by targeting a fixed, uniform distribution. MER is more "organic"—it penalizes low entropy based on the model's current state, making it more flexible and effective in practice.
- Robustness to "Label Corruption": In experiments where training labels were intentionally corrupted (wrongly labeled), MER-trained models were significantly more resilient. Because the model is discouraged from being 100% certain, it doesn't "memorize" the wrong labels as aggressively.
Key Results
The addition of MER provided consistent gains across the board:
- CTW Character Recognition: Accuracy improved from 78.53% to 79.24%. Notably, the MER-enhanced ResNet50 outperformed a much deeper ResNet152.
- ReCTS Line Recognition: Accuracy rose from 76.89% to 77.50%.
- Fine-Grained Birds (CUB): The model achieved 87.3% accuracy, outperforming several specialized architectures (like MACNN and MAMC) despite using a simpler backbone.
- Robustness: At a 20% label corruption rate on the CUB dataset, the MER model maintained 81.91% accuracy, while the standard model dropped to 79.01%.
Limitations and Future Work
The paper's theoretical derivation relies on a "strong assumption" that all negative classes are treated equally (i.e., the model is equally uncertain about all incorrect options). In reality, some characters are much more similar than others.
Future Directions: The authors suggest that further study into the regularization term could lead to even more sophisticated ways to handle class imbalance and fine-grained distinctions, potentially by incorporating character-structure knowledge (like radicals) directly into the entropy term.
Summary updated