ashishkamra79/bofa_cheques

一个包含1,000张合成银行支票图像的数据集，具有像素级准确的地面真实注释，设计用于在金融文档上对OCR和文档理解模型进行基准测试。每个样本包括一张支票图像和支票上每个文本字段的结构化地面真实数据。图像使用Google Gemini的图像生成能力生成，所有个人身份信息均使用Faker库合成生成。数据集包含100张独特的支票图像，每张复制10次并打乱以产生1,000个样本。分辨率约为960x540像素，格式为PNG。地面真实字段包括地址、分数路由号码、日期、支票号码、收款人、金额、书面金额、备忘录、签名、路由号码和账户号码。数据集还包括一个带注释的样本支票图像，用于帮助视觉语言模型将提取的字段映射到数据集中的正确地面真实列。

A dataset of 1,000 synthetic bank check images with pixel-accurate ground truth annotations, designed for benchmarking OCR and document understanding models on financial documents. Each sample consists of a check image and structured ground truth for every text field on the check. The images were generated using Google Geminis image generation capabilities, starting from a template with annotated bounding boxes. All personally identifiable information is synthetically generated using the Faker library. The dataset includes 100 unique check images, each duplicated 10 times and shuffled to produce 1,000 samples. Resolution: approximately 960x540 pixels. Format: PNG. Ground truth fields include address, fractional routing number, date, check number, payee, amount, written amount, memo, signature, routing number, and account number. The dataset also includes an annotated sample check image to help a visual language model map extracted fields to the correct ground truth columns in the dataset.