Hỏi & Đáp / AI & data

    How does image hashing catch duplicate loyalty receipts?

    Viết bởi PEKO Team.Cập nhật lần cuối: 24/05/2026.

    Perceptual hashing (pHash) produces a short fingerprint of an image that stays stable across rotation, crop, compression, and minor lighting changes. Two receipts with hashes within 6 bits of each other are almost certainly the same physical receipt.

    Xuất bản: 24/05/2026

    A determined member can rotate, crop, recompress, or photograph an already-uploaded receipt and pass any naive pixel comparison. Perceptual hashing solves this. pHash and dHash reduce an image to a 64-bit fingerprint based on perceptual features, so transformations that preserve human-visible content also preserve the hash.

    Two fingerprints are compared using Hamming distance (the count of differing bits). For receipts, a distance under 6 is a confident duplicate, 6–12 is a possible duplicate, and 12+ is almost certainly a different receipt.

    Hash on upload, not on review

    Compute the fingerprint synchronously during the upload flow. Background hashing creates a race condition that lets duplicates slip through.

    Keep a rolling window

    Compare against the last 90 days of uploads per brand. Older receipts can be archived to cold storage.

    Combine with OCR signals

    A duplicate hash plus a different extracted total is a sign of tampering, not a normal duplicate. Escalate the risk score.

    Câu hỏi thường gặp

    Why not just compare extracted totals and dates?

    Totals and dates collide all the time across legitimate receipts. Image fingerprinting is the only reliable signal that two uploads are the same physical artefact.

    Does PEKO use this?

    Yes. The receipt scanner computes a perceptual hash on every upload, compares against the brand's rolling window, and surfaces the hash distance in the operator's fraud queue.

    Tính lợi nhuận PEKO mang lại cho quán

    Bài liên quan

    People also read