Tapirscan

Find barcodes at their own angle.

Orientation-aware barcode scanning for JavaScript, Python, Rust, C, C++, Java, Swift, Android and Dart.

Documentation · Try the live demo · npm · PyPI · JavaScript · Python · How it works · Compare scanners

Tapirscan’s AI-developed algorithm finds barcode regions, estimates their orientation, and samples across the bars, including diagonally. It is built for photos and camera frames where a clean horizontal or vertical scanline may be hard to find.

Supports linear and 2D barcodes, including EAN/UPC, Code 128, QR Code, Data Matrix, PDF417 and Aztec. Retail formats (EAN-13, UPC-A, EAN-8 and UPC-E) are enabled by default; select additional formats as needed. See format coverage for all formats and variants.

• Find multiple symbols. Return decoded values and positions in the original image.
• Keep useful evidence. Inspect localized regions even when decoding fails.
• Read what the code says. Parse GS1, Wi-Fi, contact, payment and other structured payloads into named fields.
• Choose the effort. low, medium (default), high, or very-high.
• Run locally. No API key, hosted decoder, neural model, or image upload.
• Use the same core. Browser/Node WASM and native Python, Rust, C, C++, Java, Swift, Android and Dart bindings.

Supported formats

All formats below are supported in every language binding. A ✓ means the format is included in that keyword group; — means it is not included, not unsupported. common is the default.

Format identifier	retail	common1D	common	1D	2D	all	Scope
EAN13	✓	✓	✓	✓	—	✓	EAN-13
UPCA	✓	✓	✓	✓	—	✓	UPC-A; normalized from zero-prefixed EAN-13
EAN8	✓	✓	✓	✓	—	✓	EAN-8
UPCE	✓	✓	✓	✓	—	✓	UPC-E
Code128	—	✓	✓	✓	—	✓	Includes GS1-128
Code39	—	✓	✓	✓	—	✓	Standard characters; full ASCII is Code39Extended
ITF	—	✓	✓	✓	—	✓	Interleaved 2 of 5
Codabar	—	—	—	✓	—

Code32, PZN, MSIPlessey, Industrial25, Pharmacode, Code39Extended, IATA25 and Matrix25 require explicit selection; no group includes them, even all.

Python and JavaScript accept these keywords as strings, for example formats="common" in Python or { formats: "common" } in JavaScript. Native bindings provide equivalent constants. See format coverage for selection syntax, variant details and EAN/UPC supplements.

Try it

Open the camera and photo demo →

Start with Pesto, choose another example, load a photo, or scan with your camera. Medium, ZXing and ZBar are enabled initially to compare the same image. Images are processed in your browser. Choose Retail, Common or other format selections to compare the readers.

Language support

Language	Integration	Guide
JavaScript / TypeScript	Browser and Node, WASM included	JS/TS
Python	Native platform wheels; Pillow, NumPy and tensors	Python
Java	JDK 22+ JAR with a separate native library	Java
C++	Header-only C++17 wrapper and CMake installation	C++
C	Shared library with typed results and explicit buffers	C
Rust	Standalone Cargo package	Rust
Swift (iOS, macOS)	Swift package over an XCFramework; camera frames. Preview: the API may change in minor releases	Swift
Android (Kotlin, Java)	AAR with a JNI bridge; CameraX frames. Preview: the API may change in minor releases	Android
Dart / Flutter	pub.dev package for Flutter and Dart over dart:ffi. Preview: the API may change in minor releases	Dart

Every binding offers all four effort modes of the same scanner. JavaScript, Python, Rust and Dart install from npm, PyPI, crates.io and pub.dev. Java, C, C++, Swift, Android and Dart use the native library; each GitHub release attaches it prebuilt per platform, with the AAR and the XCFramework, and the guides explain how to build it from source.

Want another language? Bindings build on the C ABI; the binding checklist lists what a new one needs.

Quick start

Install Tapirscan from npm or PyPI. See the GitHub releases for release notes and downloadable packages, or build locally. Upgrading? See the migration guide.

JavaScript / TypeScript

npm install tapirscan

In a web app, tapirscan/browser/medium scans files, <img>, <video> and canvases in a worker, with no WASM setup:

import { scan } from "tapirscan/browser/medium";

const result = await scan(file); // e.g. a File from <input type="file">
console.log(result.values); // Decoded strings, e.g. ["4006381333931"]
for (const barcode of result.barcodes) {
  console.log(barcode.text, barcode.format, barcode.polygon, barcode.rect);
}

result.values contains decoded strings. result.barcodes pairs each value with its format, four polygon corners and enclosing rectangle, in input-image pixels. Both are empty when nothing is decoded. Use inspect instead of scan for timing, unread regions and diagnostics. The one-shot scan starts a worker for each call; for more than one image, create a Scanner once and reuse it.

Node and code that already has decoded pixels use the core tapirscan entry. React, Svelte and Vite setup, camera scanning and Node loading are in the JavaScript guide.

Python

Install Tapirscan and Pillow to scan a photo:

pip install tapirscan pillow

from PIL import Image
import tapirscan

with Image.open("label.jpg") as image:
    result = tapirscan.scan(image)

print(result.values)
for barcode in result.barcodes:
    print(barcode.text, barcode.format, barcode.rect)

Defaults are Medium effort and common formats. Select exactly the formats you need, for example formats=["EAN13", "QRCode"], or "1D", "2D" or "all"; each enabled 2D format adds work to every scan. Set mode="high" for more effort on difficult images. For many images, create one tapirscan.Scanner(...) and call scanner.scan.

The Python guide covers TIFF images, PyTorch tensors, and array/tensor layouts, value ranges and GPU tensors. Decoding runs on CPU.

How it differs from ZXing and ZBar

For EAN-13, the ZXing-C++ reader used by our demo primarily searches horizontal rows and, with rotation enabled, vertical rows. A tilted barcode can still work if one of those rows crosses the complete pattern.

ZBar can assemble compatible left- and right-half EAN reads from different scanlines. Tapirscan instead explicitly localizes oriented regions and samples along their geometry, with bounded retries and a small-detail recovery pass in the higher effort modes.

These are different ways of spending work on an image—not a universal speed or accuracy ranking. Read the illustrated comparison and its sources.

Choose an effort mode

Mode	When to use it
low	Live camera scanning when keeping up with incoming frames matters most.
medium	Start here: the default for photos and camera frames.
high	Difficult images when you can spend more time on each scan.
very-high	The largest effort budget, including subpixel recovery for small EAN-13 symbols.

Start with medium. Try low if scanning slows down your camera preview, or high and very-high when an image is difficult to read. Measure on your own images and devices: higher effort does not guarantee more reads on every image. Effort modes tune every format; see effort levels.

Scan fast

Speed is mostly decided by four choices: enable only the formats you need (each 2D format adds a search of the whole image), reuse one scanner, keep medium unless low reads your camera frames equally well, and pass decoded 8-bit pixels at the size you need (1-2 megapixel camera frames rather than 12-megapixel stills). On 960×960 photos, Low with retail formats takes about 5 ms per scan natively, Medium with all formats about 35-58 ms. See fast end-to-end scanning for measured costs and the fastest input path in browsers, Node, Python, Java, C, C++, Rust and mobile apps.

JavaScript also offers experimental Turbo presets for faster 1D scanning: experimentalTurbo: 2 | 4 | 8 | 16. They trade difficult-image recovery for less work. See usage and stability.

Results

Every binding returns all decoded instances, including separate labels with the same value, with polygons in source-image coordinates. result.best (every binding except C) selects the read with the highest reader-specific support. Support is a ranking heuristic, not a probability; select by format or payload when your application knows what it expects. Wrong reads, duplicates and missed symbols remain possible.

Documentation

I want to…	Start here
Integrate in a browser or Node app	JavaScript / TypeScript
Deploy on a website	Browser deployment
Switch from another scanner	ZXing · ZBar · Quagga · html5-qrcode · flutter_zxing
Scan Python images and arrays	Python
Use another native language	Rust, C, C++, Java, Swift, Android,

The demo is a separate application. Its photos, UI, and comparison engines are not included in the npm package or Python wheels.

The story behind Tapirscan

Tapirscan’s scanner algorithm was written by GPT-6 Astra, guided by my goals, experiments, and hands-on testing. I directed the development, checked results, and refined the evaluation process.

Read the development story.

Project status

Tapirscan supports the linear and 2D formats listed in format coverage. Turbo presets, the content kinds marked experimental in structured content, the JavaScript camera controller (tapirscan/camera) and the Swift, Android and Dart previews may change in minor releases. Release artifacts are published only for platforms whose release workflows pass. No general claim of superiority over ZXing or ZBar is made without a reproducible paired benchmark. See the compatibility policy for API stability.

Licensed under either the MIT License or the Apache License, Version 2.0, at your option (MIT OR Apache-2.0), copyright © 2026 Florian Nick. Third-party components retain their own licenses and notices.