What Is Image Preprocessing?

Image preprocessing is the stage of an image processing pipeline that conditions raw sensor data before analysis, converting it into a form on which measurement, recognition, or interpretation can be performed reliably. It sits between acquisition and the analysis proper, and its operations are chosen to remove or compensate for known corruptions rather than to extract meaning. Typical concerns include sensor noise, non-uniform illumination, geometric distortion introduced by optics or platform motion, and differences in scale, resolution, and dynamic range between images that will be compared. Because downstream algorithms inherit whatever defects survive this stage, preprocessing decisions often affect final accuracy more than the choice of classifier or detector that follows.

Radiometric Correction and Noise Reduction

The first group of operations acts on pixel intensities. Dark current subtraction, flat-field correction, and gain calibration compensate for the characteristics of the sensor itself, while denoising filters suppress the residual random component. Linear filters such as Gaussian smoothing reduce noise at the cost of blurring edges, so edge-preserving alternatives including the median filter, bilateral filter, and non-local means are used where boundary sharpness matters. Contrast operations form the other half: histogram equalization and its localized variant, contrast-limited adaptive histogram equalization, redistribute intensity values so that low-contrast regions become usable. Implementations of these filters and transforms are collected in general purpose toolkits such as scikit-image, which documents the algorithms alongside their parameters and assumptions.

Geometric Correction and Normalization

The second group acts on pixel positions. Lens distortion correction, rectification, resampling, and registration to a common coordinate system bring images into geometric agreement so that measurements taken from one can be compared with another. Earth observation offers the clearest example of a standardized pipeline: the Landsat Collection 2 Level-2 science products distributed by the U.S. Geological Survey apply geometric registration together with atmospheric correction, delivering surface reflectance that accounts for aerosol scattering and thin cloud so that scenes acquired months apart can be differenced. Normalization completes the stage by putting values on a common scale, whether by subtracting a per-channel mean, dividing by a standard deviation, or mapping to a fixed intensity range.

Preprocessing for Learned Models

Machine learning pipelines impose their own preprocessing conventions, driven less by physics than by the fixed input requirements of a network. The convention established during the ImageNet Large Scale Visual Recognition Challenge, in which images are resized so the shorter side reaches a set length, center-cropped to a square, and normalized by dataset statistics, remains widely used because pretrained weights assume it. Preprocessing and augmentation are distinguished by intent: preprocessing applies deterministically to training and inference alike, whereas augmentation applies random variation during training only. Mismatched preprocessing between the two phases is a common and easily overlooked source of accuracy loss, since a model receives inputs whose statistics differ from those it was fitted on.

Applications

Image preprocessing has applications in a wide range of disciplines, including:

• Medical imaging, where bias field correction and intensity normalization precede segmentation of MRI volumes
• Remote sensing and photogrammetry, where atmospheric and geometric correction enable multi-date comparison
• Biometric systems, including fingerprint ridge enhancement and face alignment before matching
• Optical character recognition, where deskewing, binarization, and noise removal precede text extraction
• Industrial machine vision, where illumination normalization stabilizes inspection under changing factory lighting
• Astronomical imaging, where flat-field and dark-frame calibration are routine steps in data reduction