Recognition Tuning

Most recognition problems come from the picture, not from the OCR engine. Work through these steps in order and check Binary View after each change. Fixing an early step often makes later tweaks unnecessary.

Video: Rock-solid readings: models, formats and filters (3:46)

On this page
  1. 1. Get a good picture
  2. 2. Fix the geometry
  3. 3. Choose the OCR model
  4. 4. Place the boxes well
  5. 5. Make the binary image clean (Tesseract)
  6. 6. Constrain what counts as valid
  7. 7. Indicator lights and dots
  8. 8. Still not reliable?
  9. Symptom quick reference

1. Get a good picture

  • Resolution: 720p is enough for most boards. What matters is the size of the digits on screen. Aim for digits at least 20–30 pixels tall. Zoom the camera in rather than cropping a tiny area of a wide shot.
  • Focus and exposure: lock both. Auto-exposure “breathing” changes the brightness of LED digits from frame to frame. Slightly under-expose LED boards so the digits don’t bloom into blobs.
  • Flicker and moiré: LED boards can flicker or show rolling bands. Match the shutter speed to the local mains frequency (1/50 or 1/100 s in 50 Hz countries, 1/60 or 1/120 s in 60 Hz countries), or defocus very slightly to remove moiré.
  • Mounting: mount the camera rigidly. If it still moves, enable Stabilize.
  • Video Settings: for USB cameras, choose a resolution and the MJPG format in Video Settings to get a clean, full-frame-rate picture.

2. Fix the geometry

  1. Rotate if the camera is mounted sideways.
  2. Crop away everything around the scoreboard.
  3. Use 4-corner if the camera looks at the board at an angle (see 4-corner correction).

Do this before placing boxes. Straight, rectangular digits are much easier to read.

3. Choose the OCR model

The OCR model list, under All boxes at the bottom of the Box tab, applies to all boxes. See OCR Models for details.

Model Best for
PP-OCR (automatic) (default) Almost every scoreboard, display and font: accurate while your computer keeps up, fast when it doesn’t
PP-OCR (accurate) / PP-OCR (fast) The same models, without switching (fast for many boxes or a slow computer)
Tesseract: General Scoreboard LED, LCD and 7-segment digits (the default before 0.3.0)
Tesseract: General Scoreboard Large Large, bold segment digits filling most of the box
Tesseract: Daktronics Daktronics-style dot-matrix/LED fonts
Tesseract: General Fonts (English) TV score bugs, apps, games and printed or on-screen text in regular fonts
Load external Tesseract model… Your own trained Tesseract .traineddata model

Start with PP-OCR: on real scoreboards it reads about three times as many values right as Tesseract with its default settings. With PP-OCR, skip step 5 (Image cleanup is for the Tesseract models). If you use Tesseract, try each of its models for a few seconds with Binary View on, and keep the one that reads most consistently.

4. Place the boxes well

  • One value per box, with a small margin. Keep frame edges, logos, colons from neighbouring clocks and other digits out of the box.
  • Set the display style to All to see the green “area read” rectangle (with Crop to text) and the blue per-character rectangles. If a character has no blue rectangle, the OCR didn’t see it.
  • If the box sometimes catches the edge of the scoreboard frame, turn up Remove specks or make the box a little smaller.

5. Make the binary image clean (Tesseract)

This step is for the Tesseract models and the dot counter: PP-OCR reads each box’s picture as it is, and its Image cleanup settings are disabled while it reads.

Select the box and look at it in Binary View (also the Binary view button in the header of Image cleanup, in the Box tab). You want solid, unbroken digits on a plain background. The settings below are in the Image cleanup section of the Box tab: click its header to open it.

What you see in Binary View Try
The whole box is black or white, or digits are lost in dark/bright areas Black & white: This box (threshold computed for this box only) or Adaptive (for uneven lighting, reflections and glare)
Digits look fine in the camera view but are broken or noisy in binary Black & white: This box, or None (the OCR then gets the grayscale image)
Small specks or dust around the digits Raise Remove specks
Segments of 7-segment/LED digits don’t touch (an “8” looks like dotted lines) Raise Thicken strokes by 1–2
Digits merge into blobs Lower Thicken strokes (0 or 1), or slightly under-expose the camera
Digits are slanted (italic) Adjust Straighten slant until they stand upright
Very tall, narrow digits Lower Squash height, or turn on Normalize digit width
Reading improves when the colours are swapped Toggle Invert colours

Binary View shows exactly what the OCR reads: Thicken strokes, Remove specks, Squash height and Straighten slant apply to the OCR with every Black & white choice. Keep Resize to 35 px on (the default). It resizes the box content to the height the OCR models were trained on. The header of Image cleanup says how many of its settings you changed, so you can see at a glance whether a box was tuned.

A box over an empty area (for example an unlit digit position, or a board that is switched off) reads as empty (∅). ScoreSight doesn’t run the OCR on a blank image, so the scoreboard models can’t invent digits there.

6. Constrain what counts as valid

Even a good picture produces an occasional wrong reading, for example while digits change or during camera flashes. These settings, in the Reading and Filtering sections of the Box tab, stop wrong readings from reaching your graphics:

  • Type: Number only allows digits, and Time allows digits, : and .. This alone removes most garbage. Use Text only for real text.

  • Format: pick a preset from the Format list (it always shows the preset in use, or Custom). A reading that doesn’t fully match the format is marked as failed filter (yellow) and is not sent as a new value. An invalid expression is shown in red: the OCR ignores it.

    Preset Accepts
    Time mm:ss.d 12:34 or 9.5 (clocks that switch to tenths in the last minute)
    Time mm:ss 12:34, 0:59
    Time ss.d 9.5, 45.1
    Time 0-59 0 to 59, with or without a leading zero (7 or 07)
    Shotclock 0-39 0 to 39, with or without a leading zero (7 or 07)
    Score 1dd 1–2 digits, or 3 digits starting with 1 (0–199)
    Score ddd 1–3 digits
    Period 1-4 a single 1–4
    Period d a single digit
    Alphanumeric letters and digits only
    Any text / Any number anything / digits only

    You can also type your own regular expression. See Format.

    Up to ScoreSight 0.3.0, Time 0-59 and Shotclock 0-39 required two digits, so a shot clock showing 9 rather than 09 was rejected and the outputs stayed at 10 for its last ten seconds. They now accept both, and boxes saved with the old presets are updated once (see One-time upgrades).

  • Remove leading zeros: turns 07 into 7. Use it for boards that zero-pad scores.
  • Smooth over last 5 readings: shows the most common of the last 5 accepted readings to smooth out flicker. It compares whole values, so a change from 12 to 9 never shows a mix such as 92. It adds a little delay when the value changes.
  • Minimum confidence: the minimum confidence of the OCR model, in percent. Readings it is less confident about are marked failed filter, and the box keeps its last good value. It is off (0) by default. With PP-OCR, 80–90 rejects most wrong readings. With Tesseract, if random wrong values still get through a strict Format, try 30–50, and lower it if good values turn yellow: Tesseract’s correct readings of real scoreboards often score only 30–50 %. See Minimum confidence.
  • Add ordinal suffix: adds st/nd/rd/th (1st, 2nd…). It is on by default for Period.

7. Indicator lights and dots

For possession arrows, bonus lights, timeout dots or balls/strikes/outs, don’t read characters. Turn on Count lights (dot counter) for the box instead. ScoreSight counts the lit blobs inside the box and reports the number. Use one box per row of lights, and use Binary View to check that each light is a separate white blob.

8. Still not reliable?

  • Try the other OCR models.
  • Make sure Detections / s isn’t higher than you need (see Performance Tuning). Fewer, cleaner readings plus Smooth over last 5 readings often beat many noisy ones.
  • Collect training data for your scoreboard and train a custom Tesseract model, then load it with Load external Tesseract model….
  • Open an issue with a short clip or screenshots (camera view and Binary View) and your exported configuration.

Symptom quick reference

Symptom Likely fix
8 read as 0, 6 or B, or 0 read as 8 Adjust Thicken strokes, try the Daktronics or Large model, improve focus
1 missing or read as 7 Widen the box slightly; Remove specks lower; check the “1” isn’t touching the box edge
First digit dropped on two-digit scores Box too tight on the left; use Score presets; turn off Skip OCR while image is unchanged to test
Letters appear in a clock Type: Time plus a Time format preset
Score shows 07 instead of 7 Remove leading zeros
Value flickers between two readings Smooth over last 5 readings, a stricter Format, better exposure
Correct readings turn yellow The Format doesn’t match (check the preset), or Minimum confidence is too high. Set it to 0 (Off) to test
A change on the board is missed Skip OCR while image is unchanged treats changes of less than 2% of the digit pixels as no change. Turn it off for that box
Box shows ∅ (empty) Nothing was read, or the box image is blank: check that the box is on the digits, then check Binary View and Invert colours
Low-contrast stadium board, dark surroundings Crop tightly, Black & white: This box or Adaptive, one box per value, Remove leading zeros, a strict Format
Readings fine in daylight, bad at night (or the reverse) Lock exposure; use This box/Adaptive black and white; save separate profiles for day and night