Pull the Tracking Number, Recipient, and Address from a Photo of a Package Label
Extract 5 fields including tracking number, recipient name, address, service class, and weight — from any carrier's package label photo, even with tape glare or curved surfaces.
5s per label · Up to 99% accuracy on printed text
What You Can Pull from a Package Label Photo
Custom Column Extraction lets you name the fields — the AI locates each one by what it means, not where it sits on the label. Even with tape glare or a curved surface.
Carrier-specific format — UPS starts with "1Z", USPS with "9400", Amazon Logistics with "TBA". The AI reads the printed number and cross-validates with the barcode.
The person or business the package is for. Usually the largest text on the label.
Street, city, state, ZIP. Sometimes partially obscured by tape glare. The AI reads through reflections by recognizing address structure rather than individual characters.
Ground, Express, Priority — often abbreviated ("GND", "PRI"). The AI reads it regardless of abbreviation style.
Package weight in lbs or kg, usually in smaller type near the bottom corner.
Tape Glare, Curved Surfaces, Dense Layouts — Why Package Labels Break Traditional OCR
A shipping label is a glossy thermal sticker taped onto a curved cardboard box, exposed to handling, light, and weather. Traditional OCR reads the reflection. Semantic extraction reads each field by what it means.
Why Traditional Tools Struggle
Tape glare and curved surfaces distort characters. A label taped over a box corner creates an angled, reflective surface. The transparent tape above the recipient address throws back light instead of letting the camera capture the text. As one developer noted in a shipping label OCR discussion, "real-world photos have glare, folds, and bad lighting" — conditions that drop conventional OCR accuracy sharply.
Dense information in a small space causes field mixing. A 4″×6″ label packs a tracking number, barcode, two address blocks, service indicator, weight, and reference fields — all separated by visual guides rather than labeled boxes. Traditional OCR dumps everything into one text block, mixing the tracking number with the ZIP code and the weight with the barcode value. Untangling takes manual work on every label.
Every carrier uses a different layout. UPS places the tracking number at the top of the label. FedEx positions it below the recipient address. USPS wraps it around the barcode. Template-based tools need a separate configuration per carrier — and a new config every time a carrier updates its label design.
How Semantic Extraction Reads Any Label
Name the columns — AI finds values by meaning, not position. Type "Tracking Number, Recipient Name, Address, Service Class, Weight". The model knows tracking number patterns (1Z..., 9400..., TBA...) and reads through glare by recognizing them — not by OCR-ing each distorted character.
Visual context resolves density. The model understands that a tracking number printed above a barcode, a recipient address block below it, and a weight value at the bottom are separate fields — even though they sit a few millimeters apart. It reads each one independently, preventing the cross-contamination that plagues traditional OCR on dense labels.
One column set works across all carriers. The same five columns extract from a UPS Ground label, a FedEx Express label, a USPS Priority label, and a DHL eCommerce label without any per-carrier configuration. Upload labels from multiple carriers into one batch — the AI processes each one independently and outputs a single spreadsheet.
From a Package Label Photo to a Spreadsheet Row in Three Steps
Photograph the Label on the Package
A package arrives at your receiving desk or doorstep. The label is taped across a box corner. The transparent tape catches the overhead light. You take a photo with your phone — JPG or PNG — capturing the full label area. Upload it into a batch. No need to peel the label off, flatten it, or remove the tape. The photo as-is is good enough to start.
Name the Fields Once
Type the column names you need: Tracking Number, Recipient Name, Address, Service Class, Weight. The same five columns work whether the label is from UPS, FedEx, USPS, or DHL. The AI processes each image independently — it finds the tracking number by its carrier pattern, reads the address block by recognizing street/city/ZIP structure, and identifies the weight by its numerical format and proximity to the unit label.
Download a Clean Spreadsheet
Processing takes about 5 seconds per label. The output is one XLSX file: each row is one photographed label, each column is one field you named. Labels from different carriers appear in the same table. Roughly 18x faster than opening each label photo, reading the tracking number, typing it in, then repeating for the address and weight (~2-3 min manual per label vs ~5s here).
When Label Photo Extraction Works Best — and When to Be Cautious
Understanding these conditions helps you get consistent results from photographed shipping labels.
When It Works Best
Well-lit front-on photos. Label photographed flat on a table or on a package face — up to 99% accuracy on printed fields.
Standard carrier labels. UPS, FedEx, USPS, DHL, and Amazon Logistics labels with printed thermal or inkjet text.
Mixed carriers in one batch. Different carriers, same column set — fields absent on a given label stay empty.
When to Be Cautious
Opaque tape covering text. Clear tape with glare is handled well; opaque packaging tape that hides the text entirely blocks extraction for the covered area.
Heavily faded thermal labels. Direct thermal labels degrade in sunlight and heat. If both the printed text and the barcode are too faded to read, extraction accuracy drops significantly.
Handwritten weight or routing annotations. Printed text extracts at 95-99%. Handwritten corrections on labels are less reliable — treat those fields as needing spot-checking.
Frequently Asked Questions
Can the AI read a Tracking Number that's partially covered by clear tape?
Yes, in most cases. The visual model recognizes a tracking number as a semantic pattern — UPS numbers start with "1Z", USPS with "9400", Amazon Logistics with "TBA". When transparent tape creates glare over part of the number, the model reconstructs it from visible characters and cross-validates against the barcode. If the tape is opaque packaging tape that completely hides the text, the AI cannot read what is not visible — but partial coverage from clear tape is handled reliably.
Each carrier formats the Service Class differently — does the tool normalize these?
The AI extracts the service indicator exactly as printed on the label: "GND" for UPS Ground, "PRI" for USPS Priority, "Express Saver" for FedEx. It reads the abbreviation or full name from wherever it appears on the label, without requiring you to know each carrier's code system. The extracted values are raw — if you need them normalized (e.g., standardising "GND" and "Ground" into a single label), a simple formula in your spreadsheet handles that in seconds after export.
Can I mix UPS, FedEx, and USPS labels in the same batch?
Yes. Define one set of column names — Tracking Number, Recipient Name, Address, Service Class, Weight — and the AI processes every label in the batch using the same definitions. A UPS Ground label and a FedEx Express label have completely different layouts, but the AI reads each one by understanding what each field means rather than where it appears on the page. The output is a single spreadsheet where labels from all carriers appear as rows with the same column structure.
What if the label photo was taken at an angle on a curved cardboard box?
Angled photos and curved surfaces reduce accuracy compared to a flat, front-on shot, but the semantic approach still outperforms traditional OCR. A 15-degree skew can cause traditional OCR to lose 30-40% accuracy because characters become distorted and line heights vary. The visual model reads through the curvature by understanding text as meaningful patterns — a name is still a name and a ZIP code is still five digits, even when they appear on a curved banner. For the best results on curved surfaces, take the photo as squarely as possible from about arm's length, capturing the full label area.
Does the AI read the barcode on the label automatically?
Yes. The model detects the barcode region, decodes the encoded value, and uses it as a cross-validation signal against the printed tracking number. On a correctly applied label, the printed number and the barcode value should match — this gives you two independent readings of the most critical field for free. When they differ (a misapplied label or a warehouse repack), the discrepancy flags that label for human review. No separate barcode scanner is needed.
Deep dives into shipping label extraction: Field-by-field accuracy breakdown for shipping labels and manifests — tracking numbers, addresses, service levels, and barcode cross-validation · How to extract tracking numbers and delivery dates from Amazon shipping screenshots across multiple carriers
Browse all photo-document capture scenarios on the Photo Documents hub page — business cards, parking tickets, prescription labels, utility meters, and more.