基於元學習的開集中文字元辨識

Abstract

近來的基於深度神經網路的模型在中文手寫辨識已高於人類辨識率。然而訓練集與真實環境的特徵分佈和類別分佈存在差距，當這類模型面對與訓練集存在特徵差異的資料準確率會下降，且無法直接用於辨識未學過的類別。因此本研究的目的是提出一個能夠在不微調或不重新訓練的情況下，模型能夠辨識不在訓練集內的類別，且對特徵變化的敏感度下降。 根據本研究的目的，我們透過訓練模型比較手寫字與印刷字相似性的方式，提出一個基於偽孿生網路架構的模型PSN-GC，透過給予新類別的印刷字範本，即可辨識不在訓練集中的類別。我們的方法相較過去研究提升了準確率，並降低記憶體用量與計算量。 實驗使用多種測試集對PSN-GC做全面的評估，測試條件可被歸類為閉集與開集。為了更進一步測試PSN-GC的極限，我們也使用甲骨文作為訓練集與測試集，因甲骨文的筆畫變化較現代手寫中文更高。以上實驗顯示我們的模型略遜於專精於閉集條件，也就是對已知類別最佳化的方法；但是與開集方法相比，我們的模型得到更高的準確率，且對特徵敏感度較低。

Recently, deep neural network-based models have achieved higher performance than humans in handwritten Chinese character recognition. However, the feature distribution and label distribution of real-world data are different from training sets. The recognition rates will drop when this type of model is evaluated on real-world data. Also, the models can not recognize unlearned categories without retraining or finetuning. Therefore, this study aims at proposing a model that can be applied to open-set and is less sensitive to feature changes. According to the purpose of this research, by training the model to compare the similarity between handwritten and printed characters, we propose a model PSN-GC based on the pseudo-Siamese network architecture. Our method improves accuracy and consumes less memory usage and computation than previous studies. The experiments use multiple testing sets to conduct a comprehensive evaluation of PSN-GC, including closed-set conditions and open-set conditions. In order to further test the limit of PSN-GC, we also use oracle bone inscriptions as the training set and testing set due to the stroke variation of oracle bone script higher than modern handwritten Chinese characters. Though our model is less accurate than the models optimized to learned categories under closed-set conditions, our model achieves higher accuracy and is less sensitive to feature changes under open-set conditions.