Still Frame From Blurry Video: Plate, Face, Text

If the camera recorded enough pixels — yes: what usually gets in the way is blur, noise and a poorly chosen frame, not the pixel count. The tool finds the sharpest frame, stacks its neighbors and removes the blur. If the characters are three or four pixels tall, the information is physically absent and no software will recover it — looking for another camera is the honest answer. A standard Russian plate is read by a dedicated neural network that answers with a list of likely plates, percentages, and a mask with “?”. In our measurements (real plates in compressed video, the plate built from neighboring frames), it reads a 48-pixel-wide plate exactly in half the cases, and the right plate is in its list three times out of four; plates 64 pixels wide or more, at least eight times out of ten. When it says “confident” (80% or more), it is right 96% of the time.

Neural upscalers paint a plausible picture, and courts have already rejected such footage as evidence: the pixels are invented. Here the main processing is classical and reproducible: frame stacking, deconvolution, contrast equalisation. Neural reconstruction exists, but as a separate button, with a warning and a label on the result. On real dashcam recordings, AI upscaling of a single frame gave 3–16% correctly read plate characters, while merging many aligned frames gave 83–86% (MF-LPR² study, 2025).

Box the plate on the frame where it is largest and press “Build the plate”. The tool finds the plate in the neighboring frames, fits the perspective for each one — the car may be approaching and turning — drops smeared and blocked frames and merges the rest on a grid 2–4 times finer. Noise averages out, and small details from different frames complete each other. The frames used are shown under the result: any of them can be taken out and the plate rebuilt.

Older cameras and DVRs write two moments in time into the even and odd lines of one frame, and store 720×576 with non-square pixels. Without field separation, sharpening turns the comb into a fence; without the pixel-aspect correction, circles become ovals and measurements lie. The tool detects interlacing itself and offers to separate the fields.

The tool does not replace an expert examination, but it provides what one values: the original frame next to the processed one, a list of steps with parameters that can be repeated, and the SHA-256 of the original file. All of it goes into a PDF report. Keep the original recording untouched — the processed frame is attached to it, not substituted for it.