synthetic-receipts-ocr
收藏资源简介:
Synthetic Receipts OCR 是一个包含 32,000 张合成热敏收据图像的数据集,涵盖美国、英国、德国、意大利和法国五个地区。每个样本包含两种图像:一张是干净的渲染图像(`image_clean`),具有热敏打印机美学风格;另一张是经过照片级真实感退化的图像(`image_photo`),模拟了透视、光照不均、阴影、热敏褪色、噪声和模糊等效果。数据集提供了精确的像素级词语边界框(`words` 和 `words_photo`)、完整的逐行文本转录(`full_text`)、从干净图像到照片图像的3x3单应性矩阵(`homography`),以及结构化的关键信息提取(KIE)字段(`fields`),包括商户、地址、电话、税号、日期、时间、收据号、商品行(名称、数量、单价、总额、税率、源产品标题)、小计、各类税费、总计、支付方式等。数据集严格遵循“字段非空当且仅当其值打印在图像上”的真实性契约。数据生成过程使用了来自亚马逊 ESCI 数据集的真实产品词汇,并强制执行地区真实性(如增值税率结构、货币/日期格式、本地化标签)。数据集分为训练集(30,000 个样本)和评估集(2,000 个样本),评估集在字体、产品标题和商户姓氏方面与训练集完全不相交,以测试模型对收据结构的泛化能力而非记忆。该数据集适用于收据 OCR、文档 KIE(如 Donut/LayoutLM 风格)、文本检测和图像去噪等任务。整个生成器代码随数据集开源,支持复现和扩展。
Synthetic Receipts OCR is a dataset containing 32,000 synthetic thermal receipt images, covering five regions: USA, UK, Germany, Italy, and France. Each sample includes two types of images: a clean rendered image (`image_clean`) with thermal printer aesthetics, and a photo-realistically degraded image (`image_photo`) simulating effects such as perspective, uneven lighting, shadows, thermal fading, noise, and blur. The dataset provides precise pixel-level word bounding boxes (`words` and `words_photo`), complete line-by-line text transcriptions (`full_text`), a 3x3 homography matrix from clean to photo images (`homography`), and structured Key Information Extraction (KIE) fields (`fields`), including merchant, address, phone, tax ID, date, time, receipt number, item lines (name, quantity, unit price, total, tax rate, source product title), subtotal, various taxes, total, payment method, etc. The dataset strictly adheres to the authenticity contract where fields are non-empty if and only if their values are printed on the image. The data generation process uses real product vocabulary from the Amazon ESCI dataset and enforces regional authenticity (e.g., VAT rate structures, currency/date formats, localized labels). The dataset is divided into a training set (30,000 samples) and an evaluation set (2,000 samples), with the evaluation set completely disjoint from the training set in terms of fonts, product titles, and merchant surnames, to test the models generalization ability on receipt structures rather than memorization. This dataset is suitable for tasks such as receipt OCR, document KIE (e.g., Donut/LayoutLM style), text detection, and image denoising. The entire generator code is open-sourced with the dataset, supporting reproducibility and extension.




