Computer vision for business image and video recognition in your workflow

Computer vision is the field in which a computer understands what appears in an image or video: which objects are there, where they are, how many and in what condition. OCR is one case of it - recognising text. Ordinary OCR returns all the text on a page as a single stream, whereas an invoice solution also has to understand the layout: which number is the amount due, which is VAT and which is the supplier number. So we combine text recognition with document layout understanding, and check the result against simple rules, such as the line items adding up to the total.

On printed documents and reasonable scans, yes, but Hebrew raises challenges worth testing up front: right-to-left text beside left-to-right numbers and dates, tables with reversed columns, stamps, and handwriting on a printed form. That is why the proof of concept runs on real documents from your suppliers. Every extracted field gets a confidence level, and a field below the threshold goes to a person for review instead of going straight into accounting. People then check only the few uncertain cases rather than retyping everything by hand.

Not necessarily. Existing security cameras are often enough for counting people or vehicles, and an office scanner or a phone is enough for documents. Quality inspection on a production line is the case that usually needs a dedicated camera and fixed lighting, because a small defect disappears in a blurred image or a shifting shadow. In the sample stage we test images from your current equipment, and if they are not sufficient we say exactly what is missing: resolution, angle, lighting or frame rate. Processing can run in the cloud, on a local server or on a small device next to the camera.

The principle is not to keep what is not needed. Counting visitors does not require knowing who came in, so processing can run close to the camera and only the number leaves it. When images or video are stored, faces and licence plates are blurred automatically before storage, and the retention period is set in advance. Amendment 13 to Israel's Privacy Protection Law broadened the obligations of anyone holding personal data, and footage of people may count as such. We do not give legal advice, but we will give your privacy adviser a full description of what is filmed, what is kept, where and why.

Fewer than people expect, because we rarely train from scratch. We start from a model that already recognises thousands of object types and adapt it to yours, so a first test can sometimes work with hundreds of labelled images rather than tens of thousands. Variety matters more than volume: images from every shift, lighting condition, angle and product type, plus enough examples of the rare case you want to catch, such as a defect. When defects really are rare, we collect them over time or add anomaly detection. Labelling is done together with you, because only you know what counts as a defect.

It will, so the design starts with which mistake costs more. In quality inspection a missed defect costs more than a false alarm, so the threshold is tuned accordingly; in product tagging it is the other way round. Every detection carries a confidence level, and borderline cases go to a person instead of triggering an action on their own. Every error that is found is kept with its image and goes into the next training round, so the model improves on exactly the cases that occur in your business. A periodic report shows how many detections there were, how many were corrected by hand and where the model struggles.