[Feature]AI removes video ads #571

Is your feature request related to a problem? Please describe.
Many third-party videos like to insert video ads, and people hope to use AI or open-source tools to remove these video ads.
Describe the solution you'd like
When playing videos from third parties, there are sudden ad insertions, which really messes up the experience.
Describe alternatives you've considered
Visual Features

Shot Boundary Features: This includes abrupt shot boundaries and gradual shot boundaries. Ad insertion points often come with shot changes. Abrupt boundaries can be identified using Manhattan distance, while gradual black frames can be detected by observing color changes. This is a typical transition method when the story moves into an ad.

Frame Content Features: This includes black frame rate, edge variation rate, and inter-frame differences. There are often black frame transitions before and after ads; the frame statistics (brightness, texture complexity) usually differ between ads and the main content. Traditional encoders can directly get these features with low computation cost.

Frame Rate Fingerprint: Main content frame rate vs. ad frame rate. In dynamic frame rate fingerprinting, the main content is usually 24 fps, while gambling-style ads are often 30 fps. Using these underlying format differences allows precise filtering.

Picture Pattern Changes: This includes frame sequence patterns (GOP structure) and forced quantization matching. Ad insertion devices may force changes in video quantization parameters or GOP structure. Detecting these encoding layer anomalies can pinpoint splice points accurately.

Special Features in HLS/SSAI Scenarios

Manifest File Structure: Including #EXT-X-DATERANGE tags and segment path features. SSAI ads leave EXT-X-DATERANGE tags or abnormal segment paths in M3U8 manifests. Recursively analyzing Blob proxies and manifest structures can detect them.

SCTE-35 Tags: This is the ad insertion signal. In broadcast-level HLS streams, SCTE-35 is the standard ad insertion signal. If it's missing, AI solutions are needed to reverse-detect ad start and end points from the content.

Segment Timing Patterns: Including segment duration and sequence number anomalies. The duration patterns and sequence continuity of ad segments may differ from the main content, so analyzing the timing patterns in the manifest can identify them.