Long-Range Barcode Scanning

BarcodeAnalyzer has a second mode for codes the native scanners give up on: across an aisle, on a high shelf, at an angle, a little out of focus. MAUI only (iOS, Android, Mac Catalyst, macOS, Windows).

<cam:BarcodeAnalyzer Mode="LongRange" AutoZoom="True" />

• Package: Shiny.Maui.Controls.Camera.Barcode, the same analyzer and package as ordinary scanning
• Engine: Shiny.Controls.BarcodeReader.Shared, dependency-free and usable without a camera
• Same results: BarcodesDetectedEventArgs, OnDetected, BarcodesDetectedCommand, Formats and ScanWindow behave exactly as in Standard mode

Why the native scanners miss distant codes

Apple Vision and Android ML Kit are handed the whole frame and shrink it before they look at it. A UPC three metres away might be 60 px wide in a 1080p frame: after that shrink its bars are thinner than a pixel, and no decoder can read bars that aren’t there. Android is worse by default, because CameraX analyzes at 640x480 unless asked otherwise.

What it does instead

Every frame:

1. Locate. A structure-tensor pass over the full-resolution luminance plane finds regions of parallel bars (1D) and of busy, direction-less texture (2D) without decoding anything. It takes a few milliseconds.
2. Read 1D codes in managed code. EAN-13, UPC-A, EAN-8, UPC-E, Code 128, Code 39 and ITF are read along a fan of up to 15 scanlines. Each line is sampled every half pixel and averaged along the bars, and edges are placed to a fraction of a pixel. Characters are matched by edge-to-similar-edge distance, which blur and ink spread cancel out of. A line that fails is retried sharpened. Lines vote, EAN digits are fused across lines, and a check-digit tie-break catches the 1/7 and 2/8 swaps that blur causes.
3. Hand 2D codes to the native decoder, enlarged. QR, Data Matrix, Aztec and PDF417 candidates, plus 1D codes the engine could not read, are cut out of the full frame, enlarged and given to Vision / ML Kit on their own. At most MaxNativeCropsPerFrame go per frame, least recently tried first.
4. Track. Codes are followed across frames. A code with a check digit reports on the first frame; one without (Code 39, short ITF) must read on UncheckedConfirmations frames. A confirmed code is not sent to the native decoder again while it stays in view.
5. Zoom (opt-in). A code that has been found but is too small to read pulls CameraView.Zoom in until it can be read. Zoom is capped so the code stays in frame, held while anything is in view, and returned once the scene has been empty for a moment. A pinch or a bound slider moving the zoom takes control back at once.

The native scanner still runs over the whole frame every FullFrameInterval frames, which catches very close codes and symbologies the locator does not frame well (Code 93, Codabar). A big, close code can therefore read a frame or two later than in Standard mode, which runs the native scanner on every frame. That is why Standard is still the default.

Usage

<cam:CameraView IsPinchToZoomEnabled="True">
    <cam:BarcodeAnalyzer Mode="LongRange"
                         AutoZoom="True"
                         Formats="Ean13,UpcA,Code128"
                         OnDetected="{Binding OnBarcodes}" />
</cam:CameraView>

public async Task<bool> OnBarcodes(BarcodesDetectedEventArgs args)
{
    foreach (var code in args.Barcodes)
        await this.Lookup(code.Value);
    return true; // keep scanning
}

As with every analyzer, results are only delivered while armed: call CameraView.Scan() (or bind ScanCommand). Return true from OnDetected to keep scanning.

From IMediaService, pass a BarcodeScanOptions (a MediaScanOptions with the barcode settings added):

var code = await media.ScanBarcodeAsync(options: new BarcodeScanOptions
{
    Mode = BarcodeScanMode.LongRange,
    AutoZoom = true
});

Analysis resolution. On Android, CameraX analyzes at 640x480 unless asked otherwise. In LongRange mode the analyzer raises its camera’s AnalysisResolution from Default to FullHd when it is attached (or when the mode is switched on). If you set AnalysisResolution yourself, it is left alone. On Apple the analyzer sees the session’s own frames, sized by VideoQuality (1080p by default).

Properties

Property	Default
Mode	Standard	LongRange turns on everything on this page (bindable).
Formats	all	Symbologies to read; comma list in XAML. Narrowing it is the biggest speed-up, and a 1D-only list skips the native crop pass entirely.
AutoZoom	false	Zoom in on codes found but too small to read (bindable).
MaxAutoZoom	4	Furthest auto-zoom goes, as a multiple of the zoom the user had.
FullFrameInterval	3	Run the native scanner on the whole frame every N frames (0 = never).
MaxNativeCropsPerFrame	2	Candidates sent to Vision / ML Kit per frame.
UncheckedConfirmations	2	Frames a code with no check digit must read on before it is reported.
ShowCandidates	true	Draw a faint box around codes found but not yet read.
BoxColor / CandidateColor	green / amber	Overlay colors.
ScanWindow	full frame	Restrict scanning to a normalized region; the locator only searches inside it.
OnDetected, BarcodesDetectedCommand, BarcodesDetected, OverlayProvider		Same as Standard mode.

Every setting except Formats and the result members is ignored in Standard mode.

BarcodeScanner.ScanLuminanceAsync(bytes, width, height) is the native-decoder entry point the analyzer uses for crops. It is public, for custom analyzers that want to hand a region to Vision / ML Kit themselves.

Measured envelope

From tests/Shiny.Controls.BarcodeReader.Benchmark (synthetic EAN-13 and Code 128 at random angles on a 1280x720 frame, noise ±4 grey levels, contrast 120; exact payload match required):

Module size	No blur	Blur σ 0.4 px	Blur σ 0.7 px
≥ 1.6 px	100%	100%	100%
1.3 px	100%	100%	66–100%
1.1 px	100%	91–100%	0%
1.0 px	~25%	~8–33%	0%

Frame time on an Apple-silicon Mac: median 3.2 ms per 720p frame for locate + 1D decode. ZXing.Net on the same frames ran at 5.4 ms and read 0–50% (its 1D readers only scan near-horizontal or near-vertical). Below about one pixel per module a single frame does not hold the code. That is the job of AutoZoom.

Not yet measured: the head-to-head against Vision / ML Kit on real devices, and real photographs. The benchmark takes a folder of photos (dotnet run -c Release -- --images <folder>, expected payload taken from the file name up to __). Those numbers are the ones to trust.

Not yet

• No Blazor version. The Blazor CameraView still uses the browser’s BarcodeDetector.
• Not yet the default mode: that waits on device timing showing it is no slower for close codes.
• Heavy defocus (blur above about 0.6× the module width) defeats the edge-based decoder. A model-based decoder that fits blurred templates to the signal is the next step.
• Codabar and Code 93 are left to the native decoder.