| 摘要: | 本研究提出CCL-OCR 模型，該模型整合了自監督對比學習框架與質心優化策略，
並融入動量更新機制。首先對發票圖像進行多樣化的數據增強，模擬真實世界的干擾因
素，以提升模型魯棒性。同時保留了動量對比學習的核心特徵，包括利用動量編碼器與
隊列機制來獲取豐富且多元的負樣本，確保特徵學習的穩定性與一致性。在此基礎上，
研究創新性地引入質心式優化損失函式，其目標是讓同一類別的樣本特徵向其質心聚集，
同時拉開不同類別質心之間的距離，從而顯著提升類內緊密度和類間區隔性。模型透過
超參數λ來平衡InfoNCE損失與質心損失的重要性。
實驗部分，CCL-OCR在八類真實發票所裁切下來的74,481張標準化字元圖片資料
集上進行實驗，涵蓋多種印表技術、字體樣式與排版差異。研究將CCL-OCR與VGG、
ResNet、SimCLR、MoCo 等模型進行比較。結果顯示，CCL-OCR表現卓越，準確率達
到96.5%，顯著超越ResNet的95.28%、MoCo的90.65%、SimCLR的85.2%及VGG的
65.22%。在精確率、召回率與F1-score三項指標上，CCL-OCR也均突破96%的門檻（精
確率96.57%、召回率96.13%、F1-score96.35%），大幅優於其他模型。這歸因於質心優
化損失有效壓縮了同類特徵的類內方差，並強化了模型對細微字元差異的辨識能力。
敏感性分析進一步探討了超參數對模型性能的影響：當λ為0.7時表現最佳，動量
對比係數接近0.7時準確率達到峰值，而溫度係數τ設為0.09時可達到最高準確率。此
外，訓練輪數增加至50輪、學習率設為0.1、批次大小擴大到256時，模型準確率皆有
提升。案例研究也顯示，CCL-OCR 在處理低對比度、輕微裁剪、筆畫繁複等複雜漢字
方面，能穩健還原結構，克服了傳統模型在細節辨識和高頻字干擾上的缺陷。;In this study, we introduce CCL-OCR, a self-supervised OCR model that unites contrastive
learning with a centroid-enhanced strategy under a momentum-update regime. First, we apply
a rich suite of data augmentations to invoice images—simulating real-world noise such as blur,
distortion, and occlusion to enhance robustness. At the same time, we preserve the strengths of
momentum contrastive learning by employing a momentum encoder and a queue mechanism
to harvest a diverse pool of negative samples, ensuring stable and consistent feature
representations. We propose a centroid-enhanced loss that pulls features of the same class
toward their centroid while pushing apart centroids of different classes. This dual objective
markedly improves intra-class compactness and inter-class separability. A single
hyperparameter λ balances the conventional InfoNCE loss against the centroid loss.
We evaluated CCL-OCR on a dataset of 74,481 standardized character images cropped from
eight categories of real invoices, covering various printing technologies, font styles, and
typesetting variations. Compared against VGG, ResNet, SimCLR, and MoCo, CCL-OCR
achieved the highest accuracy at 96.5%, outperforming ResNet (95.28%), MoCo (90.65%),
SimCLR (85.20%), and VGG (65.22%). It also surpassed 96% in precision (96.57%), recall
(96.13%), and F1-score (96.35%), significantly exceeding the benchmarks set by the other
models. These gains stem from the centroid loss’s ability to compress intra-class variance and
sharpen discrimination between visually similar characters.
Sensitivity analysis found the optimal settings to be λ = 0.7, momentum = 0.7, and τ = 0.09.
Extending training to 50 epochs with a learning rate of 0.1 and batch size of 256 further boosted
performance. Case studies confirm that CCL-OCR robustly handles low-contrast, cropped, and
complex-stroke Chinese characters, overcoming the fine-detail and high-frequency interference
issues common in traditional OCR. |
