Raw thermal counts to temperature: calibration essentials

A thermal image can look correct while carrying no usable temperature values. A video interface may deliver palette colors, contrast-adjusted grayscale, detector counts or a temperature-encoded array, sometimes through similar-looking formats. This distinction matters when you want numerical measurements rather than a picture. The practical task is to identify the output, obtain its documented calibration contract and preserve the settings needed to interpret each frame. Only then should you convert pixels into temperatures.

Identify what each pixel represents

Start with the stream definition, not the displayed image or the array's storage type. An unsigned integer array tells you how values are stored. It does not identify their physical meaning.

Some outputs described as raw have already received detector corrections. Others retain additional processing requirements. Ask which operations have been applied rather than treating raw as a precise format specification.

For display-derived data, the separate guide to mapping thermal image colors to temperature explains why a color bar cannot replace radiometric data. For a numerical workflow, prefer a manufacturer-supported measurement output whenever one is available.

Establish the stream contract before conversion

Record the acquisition path, selected stream, pixel format and processing mode. A general video adapter may expose a viewable stream without exposing the camera's measurement interface or calibration metadata.

Confirm these details against the manufacturer's documentation:

• Meaning: detector counts, apparent temperature, corrected object temperature or display brightness.
• Encoding: integer or floating point, signedness, byte order, packing, alignment and row stride.
• Valid range: usable values, invalid codes and saturation indicators.
• Geometry: image dimensions and any non-image rows or metadata fields.
• Processing: automatic contrast, rescaling, filtering, gain selection and detector correction.
• Association: how calibration information and settings relate to each frame.

A 14-bit unsigned value has 16,384 possible codes, from 0 through 16,383. Those codes do not imply any particular temperature range. They may also occupy a larger storage container, with documented alignment or packing rules.

The guide to exporting high-bit-depth grayscale images covers preserving numerical image values. Preservation is necessary, but it does not establish what those values measure. Likewise, camera bit depth describes numerical capacity, not temperature accuracy.

Obtain the calibration information that actually applies

Raw detector counts require a calibrated relationship between detector response and incident radiation or temperature. Depending on the camera, that relationship may be provided through coefficients, lookup tables, a supported software function or a documented temperature-output mode.

The required information can include:

• Camera-specific and lens-specific calibration identifiers.
• Measurement range and detector gain state.
• Detector temperature or other internal compensation inputs.
• Per-pixel correction information, where required by the documented method.
• The definition and units of every coefficient.
• Firmware or processing-version compatibility.
• Whether the conversion produces apparent temperature or scene-corrected object temperature.

Not every system exposes these items individually. A supported software interface may apply them internally and return temperatures directly. That is a valid measurement path if its inputs, assumptions and output units are documented.

Keep calibration information with the recording. A saved image can become uninterpretable if the associated range, coefficient set or correction settings are lost.

A detector non-uniformity correction reduces spatial differences between detector elements. It is not, by itself, an absolute temperature calibration. Confirm how correction events affect frame validity and whether the interface marks transitional frames.

Do not substitute coefficients from a similar camera or an unrelated example. If the supported interface does not provide the necessary conversion information, request a documented measurement output or official export method.

Use linear scaling only for a documented temperature encoding

A scale-and-offset conversion is appropriate when the output is explicitly defined as temperature-encoded data. It is not a general law for thermal detector counts.

For an output documented in kelvin, a possible encoding is:

Here, N is the decoded pixel value, a is the documented temperature increment per count and b is the documented offset. The subtraction converts kelvin to degrees Celsius. If the output is already encoded in degrees Celsius, use its specified equation instead.

Check invalid codes before applying arithmetic. Otherwise, a missing-data marker may become an apparently plausible temperature. Perform the calculation in a numerical type that preserves the required range and fractional precision.

For detector counts, the relationship may instead require a nonlinear inverse model or lookup table. Its form depends on the calibration and the signal processing already applied. A generic radiation equation cannot supply missing camera calibration coefficients.

Separate detector calibration from scene corrections

A calibrated detector measures radiation reaching the camera. Converting that radiation into an object's surface temperature also requires assumptions about the scene.

For an opaque surface, a simplified band-effective radiance model is:

The terms represent:

• L_meas: calibrated radiance reaching the camera.
• ε: effective surface emissivity.
• τ: atmospheric transmission over the viewing path.
• L_bb(T): blackbody radiance evaluated with the applicable spectral response.
• T_obj: object surface temperature.
• T_refl: reflected apparent temperature representing surrounding radiation.
• T_atm: atmospheric temperature.

This model assumes an opaque target, a simplified atmospheric path and no additional window. It is explanatory, not a substitute for the manufacturer's conversion procedure.

Rearranging gives the object's blackbody-equivalent radiance:

The final inversion uses the applicable calibrated spectral relationship. Temperature itself is not mixed linearly in this equation.

Emissivity and reflected apparent temperature become particularly significant for reflective surfaces. Atmospheric transmission can also depend on distance, humidity and wavelength. A window adds transmission, reflection and emission terms that this simplified equation omits.

Determine which corrections the camera or software has already applied. Applying emissivity correction to an output already corrected with the same assumptions introduces another error rather than improving it.

Work through a documented encoding example

Consider a hypothetical output whose documentation defines each valid count as 0.02 K, with zero offset. This example describes a temperature encoding, not an assumed property of any camera listed below.

For a decoded pixel value of 15,000:

The output increment is 0.02 K per count. That increment is quantization, not a statement that the measurement is accurate to 0.02 K. Calibration uncertainty, target properties, spatial averaging and operating conditions remain relevant.

Now suppose the same stored value is a raw detector count rather than a temperature code. The arithmetic above has no established physical meaning. You first need the applicable detector calibration and its required operating-state inputs.

This distinction also explains why fitting a convenient multiplier to a single observed temperature is unreliable. One point cannot establish both scale and offset. Even two points do not prove that a detector response is linear across the operating range.

Use reference measurements to verify the documented conversion, not to replace missing factory information with an unsupported equation. If you develop an application-specific calibration, document its restricted range, conditions, traceability and uncertainty separately.

Verify the conversion before trusting measurements

Verification should exercise the complete acquisition path, including decoding, calibration selection and scene settings. A plausible room-temperature value is not sufficient evidence.

1. Save an unchanged test frame. Preserve the original values and associated metadata before any display processing.
2. Match the official measurement settings. Use the same range, emissivity, reflected apparent temperature and other applicable inputs.
3. Compare identical pixels or regions. A cursor value and a large-area average are not equivalent measurements.
4. Check separated reference conditions. Use at least two suitable reference temperatures within the intended operating range.
5. Repeat after an operating-state change. Test relevant range transitions, correction events and acquisition restarts.
6. Record invalid and saturated pixels. They must remain excluded from temperature statistics.

For example, suppose suitable reference conditions are 27.0 °C and 60.0 °C. Your workflow reports 27.4 °C and 59.1 °C. The respective differences are +0.4 °C and -0.9 °C.

Do not average those differences and declare the conversion correct. Their different signs could indicate a scale problem, scene-setting mismatch or other measurement effect. Assess each difference against the reference uncertainty and your application's acceptance limits.

The principles in verifying per-frame exposure and gain metadata also apply here: record settings associated with the acquired frame rather than assuming a later control-panel value describes it.

Choose hardware by its documented measurement interface

Start with infrared cameras, then use the camera finder to narrow the interface and imaging requirements. Thermal imaging capability alone does not establish that a particular acquisition path exposes calibrated temperatures.

The catalog identifies LS-IRC-0044 as a thermal camera. Treat it as a candidate for technical qualification; confirm the measurement output, official software access and calibration documentation before specifying it for numerical thermometry.

The catalog also lists LS-IRC-0026 as a thermal imaging camera. Apply the same qualification process rather than inferring temperature conversion support from the product category.

For either candidate, request an example frame, its associated metadata and a documented conversion demonstration. Confirm that the proposed interface provides the required data continuously, not only through a separate still-image export.

Also separate temperature sensitivity from absolute measurement accuracy. The guide to thermal camera NETD and lens choice explains why small visible temperature differences do not establish absolute accuracy. Target size, focus and optical sampling still affect whether the intended surface dominates a measurement pixel.

Know when the usual approach is wrong

The usual preference is a supported temperature output with documented settings. Even that output can be unsuitable when its assumptions do not match your application.

A reflective target may be dominated by surrounding radiation. A small target may share pixels with the background. Saturation removes information that later calibration cannot restore. Display enhancement may improve visibility while obscuring whether the underlying measurement remains valid.

An arbitrary two-point rescaling is also a poor repair for unknown detector data. It may match the reference conditions while failing between them, outside them or after a gain-state change.

If your task only requires locating a warmer region, qualitative imaging may be sufficient. Keep that claim separate from absolute thermometry. A stable-looking image is not evidence of a calibrated temperature scale.

Final checks before implementation

Before committing to the acquisition method, confirm the following:

• The stream's pixel meaning is explicitly documented.
• Packing, signedness, byte order and invalid codes are verified.
• Calibration information applies to the camera, optics and operating mode.
• Scene corrections have a single identified owner: camera, official software or your documented processing.
• Original frames and associated metadata can be retained together.
• Reference checks cover the intended measurement range and relevant state changes.
• Missing, saturated and transitional data cannot silently enter reported statistics.

Begin the camera configuration with your measurement range, target size, viewing distance and required data interface. If the available documentation does not establish a complete conversion path, request application support before selecting hardware or writing the temperature conversion.