Computer Vision Systems • Image Preprocessing (Augmentation, Normalization)Hard⏱️ ~3 min
Domain Specific Preprocessing Constraints
Generic augmentation policies designed for natural images can break domain specific semantics, requiring careful constraint design. Medical imaging, satellite imagery, and industrial inspection have physics or regulatory requirements that constrain valid transformations.
In medical imaging, left right flips can invalidate anatomical labels. The human heart is on the left; flipping a chest X-ray horizontally creates an anatomically impossible image that teaches the model incorrect features. Similarly, arbitrary hue shifts destroy diagnostic information in histopathology, where tissue stains encode cell types through color. Hematoxylin stains nuclei blue, eosin stains cytoplasm pink; shifting hue can invert these cues. Safe augmentations include small rotations (±5 degrees to simulate patient positioning), brightness and contrast adjustments within radiometric sensor ranges, and Gaussian noise matching known signal to noise ratios (SNRs) of imaging modalities. Histogram equalization improves visibility in low contrast regions but can create false edges; use Contrast Limited Adaptive Histogram Equalization (CLAHE) with conservative clip limits around 2.0 to 3.0.
Satellite and remote sensing imagery have radiometric calibration requirements. Multi spectral sensors measure reflectance in calibrated bands; naive min max scaling or JPEG compression clips information in 12 bit or 16 bit channels. Normalization must use per band statistics over the sensor's dynamic range, not 8 bit assumptions. Geometric augmentations are also constrained; geospatial coordinates and orientation matter for tasks like change detection or infrastructure mapping. Random rotations can break cardinal direction features. Photometric augmentations should simulate atmospheric scattering and seasonal lighting variations, not arbitrary color shifts.
Industrial defect inspection often relies on color or texture cues that encode material properties. A discoloration on a semiconductor wafer indicates contamination; hue shifts can hide the defect. High dynamic range (HDR) imaging in low light conditions exposes another failure mode: histogram equalization amplifies sensor noise and creates halo artifacts. For temporal data like video surveillance, applying independent augmentations per frame destroys motion cues. Always apply consistent transforms across the temporal window.
The lesson is clear: understand the domain physics and label semantics before enabling augmentation families. Start conservative, measure per class metrics, and expand gradually.
💡 Key Takeaways
•Medical imaging: left right flips break anatomical labels (heart location), hue shifts destroy histopathology stains (blue nuclei, pink cytoplasm)
•Satellite imagery: 12 bit or 16 bit radiometric calibration requires per band normalization over sensor dynamic range, not 8 bit min max scaling
•Random rotations in geospatial tasks break cardinal direction features; change detection and infrastructure mapping need orientation preserving transforms
•Industrial defect inspection: color discoloration encodes material defects; arbitrary hue shifts can hide semiconductor contamination or coating flaws
•High dynamic range (HDR) and low light: histogram equalization amplifies sensor noise and creates halo artifacts in dark regions, use CLAHE with clip limit 2.0 to 3.0
📌 Examples
Meta medical AI: disables horizontal flips for chest X-rays after anatomical label errors, allows ±5 degree rotation and brightness ±0.2 only
Google Earth Engine models: normalize Sentinel 2 satellite bands using per band statistics over 10000 reflectance range, avoiding JPEG compression artifacts
Tesla dashcam models: apply consistent brightness and motion blur across 10 frame clips to preserve temporal coherence for trajectory prediction