Our "Business Card Scanner" solution aims to automate the extraction and validation of information from business card images. Using state-of-the-art Optical Character Recognition (OCR) and Natural Language Processing (NLP) methods, the data contained on the business cards is captured, validated and converted into a structured JSON format. This solution offers an efficient and reliable way of extracting important information such as names, positions, company details and contact data from various card formats. We attach great importance to data protection - the Business Card Scanner can be run locally and the data remains on the executing device.

Business cards come in a wide variety of designs, fonts and languages, which makes the manual processing and recording of information time-consuming and error-prone. The biggest challenges include

• Versatile formats: Different layouts, fonts and logos require flexible recognition methods.
• Special characters: Telephone numbers, e-mail addresses and URLs often contain special characters that need to be recognized precisely.
• High accuracy: The extracted information must be error-free and well-structured to enable direct further processing.

Our solution combines powerful OCR models and a locally run large language model (LLM) to ensure accurate and structured extraction of business card data. The process includes the following steps:

1. Text extraction with OCR:

• Use of the EasyOCR and PyTesseract OCR libraries for reliable recognition of text on business cards.
• EasyOCR offers excellent text detection, while PyTesseract excels at recognizing special characters, such as in phone numbers and email addresses.
• A robust extraction is achieved by combining both models.

2. Data validation with LLM:

• Use of the Phi3.5 Language Model to validate the extracted data and correct any inaccuracies in the OCR.
• The model ensures that the results are contextually and structurally correct.

3. Structured output in JSON

• The validated data is stored in JSON format, which includes fields such as first name, last name, role, company name, address, telephone, mobile number, fax, e-mail and website.

The prototype developed provides a flexible and scalable solution for extracting information from business cards. The most important results include

• Precise text extraction: The combination of EasyOCR and PyTesseract ensures a high degree of accuracy in text recognition.
• Context-related validation: The use of the Phi3.5 LLM ensures complete and reliable data.
• Structured output: The data is provided in a standardized JSON format that enables easy further processing.
• Broad applicability: The solution can handle a variety of fonts, layouts and languages.