If you have ever zoomed in on a frame and found the subject fuzzy or out of focus, you’ve met image blurriness. In computer-vision pipelines,
blur is a dataset quality issue that reduces information content, confuses feature extractors, and ultimately lowers model accuracy in detection tasks. Voxel51’s FiftyOne includes built-in
blurriness scanning that automatically flags overly blurry (or abnormally sharp) samples so that teams can triage them in seconds.
What is image blurriness in computer vision?
Image blurriness is the loss of high-frequency detail that makes edges crisp and textures distinct. It can come from hardware factors (out-of-focus lenses, long exposures, small apertures), scene dynamics (fast motion, atmospheric haze), or post-processing (improper resizing or heavy compression). Models struggle to learn or identify objects when pixels are not clearly distinguished from each other.
Why blur hurts model performance
Deep learning models learn by looking at edge contrast and fine textures. When those cues disappear, classifiers mislabel whole scenes and detectors miss small objects. Studies show that a few percent of blurry training samples can lower mean Average Precision (mAP) by double digits on benchmarks such as COCO or Waymo Open.
What are common sources of blur?
Defocus blur
- Tell-tale signs: the entire frame looks soft; when you zoom in, points appear as “circles of confusion.”
- Typical fixes: re-capture with the correct focus distance, enable or recalibrate autofocus checks.
Motion blur
- Tell-tale signs: linear streaks in the direction of camera or subject motion; duplicated “ghost” edges.
- Typical fixes: use a faster shutter speed, switch to a global (not rolling) shutter, or stabilize the rig.
Compression blur / resizing artifacts
- Tell-tale signs: blockiness and “mosquito” noise, especially around high-contrast edges, text, or logos.
- Typical fixes: re-encode with less aggressive compression settings and preserve original-resolution images.
How to measure blurriness
Variance of Laplacian
The quickest check is the variance-of-Laplacian test. A Laplacian filter highlights edges; then you calculate the variance (how spread-out the pixel values are). A sharp photo has a high variance because it contains many strong edges, while a blurry one has a low value. In practice you pick a threshold, and everything below it is marked “too blurry.” This test needs only a few lines of code and runs fast on large datasets.
Fast Fourier Transform (FFT)
A more detailed method looks at the picture in the frequency domain with a Fast Fourier Transform (FFT). After an FFT, sharp edges appear as strong high-frequency components. If those components are missing—or very weak—the image is probably out of focus or motion-blurred. You can quantify this by summing the energy in high-frequency bins and comparing it to a threshold.
MTF50
For camera‐hardware work you’ll often see MTF50 (Modulation Transfer Function at 50 % contrast). It tells you the spatial frequency where the lens-sensor system cuts contrast in half. Higher MTF50 values mean the system keeps fine detail better. Because it models the optics themselves, MTF50 is standard in lens quality reports and test charts.
How to detect and filter blurry images with FiftyOne
FiftyOne makes it easy to spot and act on blurry images. Run a Blurriness scan in the
Data Quality panel and each sample is assigned a numeric blurriness score; the panel shows an interactive histogram where you drag a slider to set a threshold, and the sample grid instantly filters to only the frames that cross it. You can then add tags such as “blurry–review” or “remove” to those images, or save the threshold so future scans flag new problems the same way.
From dataset to deployment
Keeping datasets sharp pays dividends at inference time—especially in safety-critical domains like autonomous driving or medical imaging. By integrating blur detection directly into your curation workflow, FiftyOne helps you ship models that stay focused in the real world..