Pixelation - definition
Pixelation is an image transformation that replaces a selected area with larger, uniformly colored blocks. In image and video anonymization, the selected area is usually a detected face or license plate. The transformation reduces local spatial detail by representing multiple original pixels with one output value.
Pixelation is often called mosaic masking. It differs from Gaussian blur, which spreads pixel values across neighboring pixels using a mathematical filter. Pixelation preserves block boundaries and large color regions, while blur produces gradual transitions. Neither method is automatically equivalent to anonymization. Whether a pixelated face or license plate can still be linked to a person depends on the source resolution, block size, viewing distance, surrounding context, and the capabilities of a potential recipient.
A simplified pixelation operation can be described as follows:
Ip(x,y) = mean(I(u,v))
Here, I(u,v) represents the original pixels within a block, and Ip(x,y) assigns the same average color value to every output pixel in that block. Some implementations use the median, a sampled pixel, or another representative value instead of the mean.
How pixelation is applied to faces and license plates
Pixelation is normally applied after the relevant region has been identified. In an automated video anonymization workflow, face detection identifies the location of a face in each frame. A separate license plate detection process identifies license plates. The system then applies a pixelation mask to the selected bounding box or polygonal region.
A reliable implementation must account for motion, partial occlusion, camera angle, and changes in object size. In video, a mask must also remain aligned over time. If tracking fails between detected frames, parts of a face or license plate can become visible.
- Face detection: locates faces without determining a person’s identity.
- License plate detection: locates the plate area before a masking transformation is applied.
- Tracking: associates a detected face or plate across consecutive video frames.
- Mask expansion: adds a margin around the detected region to reduce the risk of exposed edges.
- Manual review: checks missed detections, incorrect masks, and areas where contextual information remains identifying.
Gallio PRO uses face detection and license plate detection for automatic masking of faces and license plates in imported images and video files. It does not perform real-time anonymization or video stream anonymization. Company logos, tattoos, name badges, documents, and content shown on monitors are not automatically detected. These elements can require manual masking in the built-in editor.
Key pixelation parameters and quality metrics
Pixelation strength is primarily determined by the relationship between the block size and the size of the protected object. A block size that is sufficient for a distant face may be inadequate for a close-up face recorded at a higher resolution.
Parameter or metric | Technical meaning | Privacy relevance
|
|---|---|---|
Block size | The width and height of each mosaic block in pixels. | Larger blocks remove more detail but can obscure more of the surrounding image. |
Face or plate width | The size of the detected object in the original frame. | Mask settings should be assessed relative to object size, not only to full-frame resolution. |
Mask margin | Extra area added outside the detected bounding box. | Reduces exposure of facial contours, plate characters, and detection-boundary errors. |
Detection recall | The proportion of relevant faces or plates that the system detects. | Low recall creates unmasked personal data in output files. |
False positive rate | The proportion of non-target objects masked incorrectly. | Excessive masking can reduce evidentiary or operational value. |
Temporal stability | Consistency of mask placement between video frames. | Unstable masks can reveal protected information during motion. |
Testing should use representative footage, including low light, motion blur, profile views, reflections, crowded scenes, and partially obscured license plates. A privacy review should assess both the masked object and the wider scene. Clothing, location, vehicle markings, voice, timestamps, and sequence context can still enable identification even when a face is pixelated.
Pixelation, blur, and resistance to reidentification
Pixelation and blur are visual obfuscation methods, not access-control mechanisms. They alter visible information in a released image or video. They do not prevent an authorized recipient from combining the output with other information.
Research has shown that machine learning can infer information from images that appear difficult for humans to interpret. McPherson, Shokri, and Shmatikov demonstrated that deep learning methods could defeat several image obfuscation approaches, including pixelation, under experimental conditions ( Defeating Image Obfuscation with Deep Learning, 2016). The result does not mean that every pixelated image is identifiable. It shows that visual appearance alone is not a sufficient privacy test.
Pixelation may be less resistant than strong blur when the block pattern preserves distinctive facial structure, such as eye position, hairline shape, or plate-character layout. Conversely, a weak blur can retain readable or recognizable features. There is no universal block size, blur radius, or compression setting that guarantees resistance to reidentification for every image.
Practical controls for pixelated footage
Organizations should treat pixelation as one control within a documented image and video disclosure process. The appropriate setting depends on the intended recipient, permitted use, image quality, retention period, and the consequences of identification.
- Define the release purpose and identify which faces and license plates require masking.
- Test pixelation settings on footage that represents the highest-risk recording conditions.
- Review samples at native resolution, not only in a reduced preview.
- Check consecutive video frames for mask drift and brief exposures.
- Assess contextual identifiers outside the pixelated region.
- Record the masking method, software version, review steps, and approval decision.
Standards and references
No International Organization for Standardization (ISO) or National Institute of Standards and Technology (NIST) standard defines a universal pixelation threshold that makes faces or license plates anonymous. The risk assessment must be specific to the dataset and disclosure context.
- ISO/IEC 29100:2011, Information technology - Security techniques - Privacy framework. This standard describes privacy principles and terminology relevant to privacy controls.
- NISTIR 8053, De-Identification of Personal Information, 2015. This publication explains that de-identification risk depends on data, recipients, and available auxiliary information.
- McPherson, Shokri, and Shmatikov, Defeating Image Obfuscation with Deep Learning, 2016. This paper evaluates attacks against several visual obfuscation methods.
- Newton et al., Preserving Privacy by De-Identifying Face Images, IEEE Transactions on Knowledge and Data Engineering, 2005. This paper discusses technical approaches to face-image de-identification.