Sokheng/khmer-synthetic-ocr-v1-100k

Khmer Synthetic OCR Dataset v1 是一个包含100,000个样本的合成高棉语OCR训练数据集，专注于收据、菜单、价格标签和标牌等领域，这些领域通常缺乏真实标注数据。数据集分为训练集（95,000个样本）和测试集（5,000个样本）。内容方面，70%为单行文本（包括单词、短语、价格以及高棉语和英语混合内容），30%为多行收据块（包含完整收据的标题、项目、总计和分隔符）。标签类型多样，例如高棉语单词（如“ភ្នំពេញ”）、短语加价格（如“អង្ករ ៥,០០០៛”）、拉丁语混合（如“Total $12.50”）、收据块（包含多行文本）、日期/时间和地址。数据生成使用Pango + Cairo + HarfBuzz渲染器，支持完整OpenType字形处理（确保正确的高棉语连字），采用49种高棉语字体（如Battambang、Noto Sans Khmer等），字体大小为20-40 pt，背景包括真实纸张纹理照片、纯色和高斯噪声，颜色为深色墨水变体（黑色、蓝色、红色、绿色、棕色），单行文本有±3°的倾斜。文本标签来源于两个公开高棉语数据集：seanghay/khmer-dictionary-44k（提供单词和短语）和KiteAether/c_khmer_gemma_ocr_train（提供短词和短语级标签）。数据集适用于图像到文本任务，许可证为CC-BY-4.0，允许研究和商业用途，但需注明出处。

Khmer Synthetic OCR Dataset v1 is a synthetically generated dataset containing 100,000 samples for Khmer OCR training, focused on receipts, menus, price tags, and signage — domains where real labeled data is scarce. The dataset is split into a training set (95,000 samples) and a test set (5,000 samples). It consists of 70% single-line text (including words, phrases, prices, and mixed Khmer+English content) and 30% multi-line receipt blocks (with full receipts containing headers, items, totals, and dividers). Label types include Khmer words (e.g., ភ្នំពេញ), phrases with prices (e.g., អង្ករ ៥,០០០៛), Latin mix (e.g., Total $12.50), receipt blocks (multi-line text), dates/times, and addresses. The data is generated using a Pango + Cairo + HarfBuzz renderer with full OpenType shaping for correct Khmer coeng ligatures, employing 49 Khmer TTF fonts (such as Battambang, Noto Sans Khmer, etc.), font sizes of 20–40 pt, backgrounds of real paper/texture photos, plain colors, and Gaussian noise, colors in dark ink variants (black, blue, red, green, brown), and skew of ±3° for single-line text only. The word pool is derived from two public Khmer datasets: seanghay/khmer-dictionary-44k (for single words and example phrases) and KiteAether/c_khmer_gemma_ocr_train (for short word tokens and phrase-level labels). The dataset is designed for image-to-text tasks, licensed under CC-BY-4.0, and free for research and commercial use with attribution.