Detect Looping Video Frames with Perceptual Hashing in Python

Plan

Write a program to detect loops in a video using Python, despite the author's limited experience with video processing.

First Attempt

Viewing a video is like rapidly flipping through images; each frame is an image array of RGB values. By counting identical frames with two dictionaries—one for seen frames and another for duplicate frames—the author attempts to spot repeats.

Code (shown as an image):

The script ran for about an hour on a MacBook Pro, producing results that highlighted matching frame pairs such as 5928 with 2048454 and 5936 with 2048462.

Complexity Increases

The program identifies identical frames to determine looping, but some visually identical frames were not flagged due to compression noise. Subtracting one frame from another revealed minor pixel differences caused by video compression.

Hashing each image converts it to an integer; identical images yield the same hash, but the goal is to have similar hashes for slightly different frames as well.

Simplifying the Compression Issue

Perceptual hashing (aHash) produces similar hashes for visually similar inputs, useful for reverse image search. For video, the hash function must be tolerant to compression noise yet sensitive enough to distinguish adjacent frames.

Be tolerant: frames differing only due to compression should hash to the same value.

Be sensitive: adjacent frames should produce different hashes.

Choosing aHash Parameters

The author experiments with resolutions 8×8 and 64×64. The 8×8 resolution loses too much detail, while 64×64 retains more information but may still miss subtle differences.

Testing various resolutions on similar videos yields match buckets such as [8,108], [9,109], and [10,11,110,111], indicating occasional false positives.

Metrics considered include the number of matching buckets, average frames per bucket, and maximum bucket size, aiming for many small buckets (ideally two frames per bucket for a single loop).

Resolution 24 provides a better balance, reducing excessive matches.

Results

Replacing the original hash with the new aHash produced many matching frames. Sample matching frame indices include 4262, 72096, 124855, 132392, 147466, 162540, 170077, 185151, 207762, 252984, etc.

Converted to timestamps (seconds), some matches occur at 142.07 s, 2403.2 s, 4161.83 s, 4413.07 s, 4915.53 s, 5418.0 s, 5669.23 s, 6171.7 s, 6925.4 s, and so on, often aligning with the midpoint of the slapping action in the video.

These findings suggest the video contains repeated segments, supporting the hypothesis that the 24‑hour slapping video is fabricated.

Author: Python developer Source: http://developer.51cto.com/art/201706/543686.htm