Image preprocessing: point, local and global operations

Image preprocessing is easier to specify when you separate two questions: which pixels determine the processing parameters, and which pixels supply values during application. This distinction matters when an inspection changes after a background, crop or exposure adjustment. For inspection developers and system integrators, the practical goal is repeatable behavior, not a disputed label. You can describe each operation precisely, predict its sensitivity to scene changes, and test it before committing to a threshold or measurement.

Separate parameter estimation from pixel application

A useful specification gives each operation two scopes. The estimation scope identifies the data used to calculate parameters. The application scope identifies the input samples read when producing an output pixel, once those parameters are available.

For a pointwise mapping, write:

Here, I is the input image, J is the output, p is a pixel position, and S is the estimation region. The parameter set theta might contain a threshold, intensity limits, or a lookup table.

With parameters held fixed, calculating J(p) reads only I(p). That makes the application pointwise. However, if the parameters are recalculated from the complete image, changing a distant pixel can alter J(p). The complete procedure therefore has global dependence.

This is why histogram equalization can accurately be described as a pointwise mapping derived from global statistics. Neither description is sufficient alone when specifying inspection behavior.

Pointwise processing is also a limiting case of neighborhood processing, with a neighborhood containing only the current pixel. For documentation, keeping pointwise operations separate usually makes dependencies clearer.

Compare operations using both scopes

The following table describes common implementations. The parameter source matters as much as the operation name.

A spatially invariant rule is not necessarily global. A fixed threshold uses the same rule everywhere, yet it never needs another pixel's value. Similarly, a mean filter can use identical weights throughout an image while remaining local.

Record scope explicitly rather than relying on a label such as “normalization.” A requirement like “estimate limits from the inspection region on every frame, then apply the mapping to that region” is much easier to reproduce.

Understand normalization and histogram equalization

Normalization depends on where its parameters come from

Min–max normalization maps an input interval into an output interval. For an illustrative 8-bit output:

The limits a and b could be fixed calibration values, per-image extrema, or selected percentiles. Those choices produce different behavior even though the application formula stays unchanged.

Per-image extrema make the complete operation globally dependent. Fixed limits make it pointwise during production. Percentile limits reduce sensitivity to isolated extremes, but still depend on the intensity population used to estimate them.

If b = a, the expression is undefined. Specify a constant-image policy, such as returning a defined value and flagging insufficient contrast. Do not leave this case to an undocumented library default.

Mean-and-standard-deviation normalization has the same distinction. Parameters computed from each image create image-wide dependence. Parameters stored from a training dataset remain fixed during inference. The training pipeline input checks explain why parameter values, channel order and scaling must agree with the deployed model.

Equalization builds a mapping from an intensity distribution

A simple global equalization convention uses the cumulative distribution function, or CDF:

Here, N is the number of pixels used to estimate the histogram, and L is the number of output levels. Some implementations subtract the first nonzero CDF value and rescale. State the convention when exact numerical agreement matters.

Suppose an image contains equal populations at intensities 50, 100, 150 and 200. At intensity 100, the CDF is 0.5. With 256 output levels, the stated convention gives round(255 × 0.5) = 128.

The lookup reads one input value per output pixel, but the lookup table depends on the full estimation population. Changing the background can therefore change the output assigned to an unchanged part pixel. Equalization also does not guarantee a perfectly uniform output histogram for discrete images.

Work through a scene-change example

Consider a grayscale inspection using per-image min–max normalization. A target pixel remains at intensity 120 between captures. Initially, the estimated limits are 100 and 200:

A bright object then enters the estimation region. The minimum remains 100, but the maximum becomes 255:

A downstream rule accepting values of at least 40 now changes its decision, despite the target pixel remaining unchanged. This is a dependency problem, not necessarily a camera fault.

You have several possible fixes:

• Store validated limits when absolute intensity carries inspection meaning.
• Estimate parameters from a stable reference region rather than the changing background.
• Use percentile limits when isolated extremes are the main problem, then validate population changes separately.
• Replace the downstream threshold with a feature whose behavior matches the intended normalization.

A stable reference region must actually remain representative. A reference patch that becomes obscured can make every output pixel wrong together.

For a controlled before-and-after check, distinguish histogram changes from spatial changes using the image comparison methods. Similar histograms do not establish that corresponding pixels or defects are unchanged.

Choose neighborhood processing for the right reason

Neighborhood processing uses surrounding samples to calculate an output. A linear filter forms a weighted sum:

A mean filter uses equal weights that sum to one. A median filter selects the middle-ranked value and is nonlinear; it is not a convolution with fixed weights.

For a 3 × 3 neighborhood containing eight values of 100 and a center value of 10, the mean is (8 × 100 + 10) / 9 = 90. The median is 100. Both suppress the isolated dark sample, but they produce different outputs and behave differently near edges.

That dark sample might be noise, or it might be the smallest defect you need to detect. Choose the window against the minimum relevant feature, not just the apparent smoothness of the output.

Also specify border handling. Padding with zeros can create artificial dark borders; replication and reflection create different edge neighborhoods. For measurements near an image boundary, test the exact implementation.

The neighborhood entropy map explanation extends this distinction to a local statistical feature. Each output describes a surrounding region rather than merely remapping its center intensity.

Adaptive equalization needs additional care: tile histograms and interpolation determine its effective spatial dependence. Record tile geometry, clipping settings and interpolation behavior instead of calling the whole method simply local.

Select a predictable inspection pipeline

Start from the property the inspection must preserve. Absolute intensity, relative contrast, edge position and texture statistics do not respond identically to preprocessing.

Keep the display branch separate from the measurement branch when their purposes differ. Automatic contrast can improve visibility without being suitable for quantitative inspection.

When a thresholded output is the inspection feature, use a live binary inspection preview to examine boundaries, missing regions and noise. Display the raw input alongside it so that an apparent segmentation improvement can be traced to its cause.

The usual preference for fixed mappings is wrong when the task deliberately requires adaptation to valid contrast changes. In that case, define the estimation region and adaptation behavior as part of the inspection specification. Validate changes in part coverage, background, glare and empty frames rather than testing only nominal images.

Keep acquisition and deployment separate from algorithm scope

Choose acquisition hardware around the image requirements first. The area scan cameras category and camera finder support that selection; a preprocessing label does not determine the required camera interface.

The LS-CMR-0053 is an acquisition candidate to evaluate against your capture requirements.

The LS-CMR-0166 is another acquisition candidate. Neither choice establishes whether your software estimates parameters globally or applies a fixed mapping.

For embedded deployment, review vision controllers against the actual runtime, image buffers and processing workload. Confirm software compatibility and licensing separately; do not infer algorithm support from a controller category.

Measure latency using representative images and the complete pipeline. Global histogram operations generally require an estimation pass before mapping, while neighborhood operations need access to surrounding samples. Buffering, memory traffic and implementation details can matter more than the scope label alone.

Before selecting hardware, document image dimensions, pixel format, capture rate, latency limit and required intermediate precision. Then use the camera configuration starting point to organize the acquisition requirements.

Verify dependencies before release

Use a small, repeatable test set to expose unintended coupling:

1. Hold the target pixels constant. Change only the background and compare the target output values.
2. Change the estimation region. Confirm that cropping or masking has the documented effect on parameter calculation.
3. Test a constant image. Check normalization denominators, empty masks and low-contrast policies.
4. Insert isolated extremes. Separate sensitivity to outliers from sensitivity to broad population changes.
5. Test small defects near borders. Verify neighborhood size, padding and feature survival.
6. Save parameters with outputs. Record limits, thresholds or lookup-table identifiers alongside the processing configuration.
7. Repeat downstream decisions. Compare inspection outcomes, not only processed-image similarity.

A defensible specification names the parameter source, update frequency, application neighborhood, numeric precision and failure policy. That description remains useful even when different software libraries use different classification terms.

If a pipeline changes decisions when only unrelated scene content changes, establish the intended dependency before replacing hardware. Bring representative raw images and the acceptance criteria to applications support to review the acquisition and processing requirements together.