FreeBarcode Scanner

What this barcode scanner does

It decodes 1D and 2D barcodes from a live camera, from JPG/PNG/GIF/BMP/WebP images, from MP4/WebM/MOV video, or from every page of a PDF or multi-page TIFF — there is no button to press, detection starts on its own. Each hit is drawn on the frame as a bounding box. Benchmark mode scores a whole image set against ground truth: recall, precision, symbology agreement, box overlap and per-page timing.

Worked example: scoring a dataset you can rebuild yourself

Start with the small synthetic fixture shipped beside this page. Download sample-qr.png, choose File (image or video) and open it in Default mode. The page detects the symbol without a button press and returns one result: [QR_CODE] https://www.dynamsoft.com/codepool/demos/barcode-scanner/. Expressed as structured output, the decode is:

{
  "count": 1,
  "results": [
    {
      "format": "QR_CODE",
      "text": "https://www.dynamsoft.com/codepool/demos/barcode-scanner/"
    }
  ]
}

For a larger test, the barcode generator exports a dataset as one ZIP — images/*.png plus annotations.json — and this page reads that archive directly. Drop the ZIP on the upload zone in Benchmark mode and every image arrives with its expected values already attached, including the degradation each image was built with. Ground truth is matched on text, and the other two fields are optional:

{
  "images": [
    {
      "file": "sample1.jpg",
      "barcodes": [
        {
          "text": "9781234567897",
          "format": "EAN_13",
          "points": [[10, 20], [310, 20], [310, 140], [10, 140]]
        }
      ],
      "degradations": { "blur": 1.2, "rotate": -7 }
    }
  ]
}

Four columns come back for each row, and they do not all mean the same kind of thing:

Column behaviour as implemented in this page, not generic definitions.

Column	What it counts
Missed	Expected values that never came back. Listed under Details beside the capture conditions, so a miss can be read against the blur or rotation that caused it.
Symbology	How many correct values came back under the annotated symbology — informational. A correct value is never downgraded because the reader named the format differently.
Box IoU	How many decoded boxes overlap the annotated box at IoU ≥ 0.5, plus the mean — also informational.
Unmatched decodes	Values that read but matched nothing expected. Either a wrong read or a barcode the annotations do not cover; the page does not guess which.

Known naming differences are folded together before comparing — CODE_39_EXTENDED with CODE_39, UPC_A with an EAN_13 that starts with 0, MSI with MSI_CODE. Anything else is reported rather than normalised away. If you have no dataset of your own, Dynamsoft's challenging-images set ships 68 photos and 514 annotated barcodes in exactly this format.

Failure case: a "miss" that is really a setting

The stock ReadBarcodes_Default preset enables 32 of the 52 formats the SDK knows. It leaves out DotCode, both Pharmacode tracks and the postal symbologies — and the barcode generator produces DotCode and Pharmacode. Benchmark a dataset containing them against the preset alone and the report shows misses that are not reading failures: the reader was never asked to look. This page sets those three bits on top of the preset through the simplified settings API and re-applies them after any settings reset. Measured on bundle 11.6.3200, a clean DotCode and a clean one-track Pharmacode both decode once the bits are set, and neither decodes without them.

Supported and not supported

Capability	This tool
Symbologies	1D and 2D, with the preset's 32 formats plus DotCode and both Pharmacode tracks enabled. A custom JSON template can change the list entirely.
Camera	Live scanning with format selection. Requires HTTPS or localhost and camera permission.
Images	JPG, PNG, GIF, BMP, WebP, single or multi-select, plus clipboard paste.
Video	MP4, WebM and MOV, scanned frame by frame as the file plays. Plus a clip recorder that replays the live stream through the same decoder.
PDF and TIFF	Decoded page by page, up to 50 pages, rasterized at up to 2000 px on the long edge. PDF goes through pdf.js; TIFF through UTIF.
TIFF compression	Raw, G3, G4, LZW, JPEG, Deflate (including Adobe Deflate), PackBits, Thunder and Nikon. Other codes are rejected with a message rather than decoded to noise.
Large files	Documents are capped at 50 pages and rasterized to a 2000 px long edge; images are decoded in the tab, so browser memory is the practical file-size ceiling and nothing is uploaded.
Benchmark and ground truth	Yes, with an HTML report export. Scoring needs annotations.json; without it you still get counts and timing.
Reading printed text	Not offered — this is a barcode decoder. A VIN stamped into metal, or an expiry date printed next to the symbol, is not read.
Server-side batch processing	Not offered. Everything runs in the tab; there is no upload endpoint and no CLI.
Privacy and network	Images, video and documents are decoded on the device and are not uploaded. The page does fetch the SDK bundle, the licence, pdf.js/UTIF on first use, and demo analytics.

Use the tool, or build it into your app

Limits worth knowing before you rely on it

Why is a clean DotCode reported as missed?

Most likely a custom template is loaded. The three extra format bits are applied on activation and after a settings reset, but a template you import replaces the format list wholesale. Re-enable DotCode and Pharmacode in that template, or reset to the default.

Why does a barcode read from the camera but not from my video file?

The two paths differ in what reaches the decoder. Video is sampled per frame, so a code that is only sharp for a few frames can fall between samples, and downscaling in the file itself loses module size. Record the live stream with the clip recorder and scan that clip: if the clip decodes, the live miss was timing rather than the frames.

Is the recall figure a production accuracy number?

No. It describes this reader on this image set with these annotations. A different camera, lens distance, print quality or lighting will produce a different number, and a dataset you built with the generator is cleaner than a real warehouse.

Why does a 60-page PDF stop at page 50?

The page cap is deliberate: each page is rasterized to up to 2000 px and held for navigation, so an unbounded document grows memory without limit. Split the document, or scan the pages you need.