Comparison of Cropping Resistance
pHash is based on global low-frequency DCT features. Local cropping changes the overall spectral distribution and often causes significant changes in the hash, so its robustness against “local missing content” is relatively weak. For this reason, academia has proposed improvements aimed at crop/rotation robustness to compensate for the shortcomings of traditional global hashing.
PDQ (images) is more easily affected by “content removal/occlusion” modifications. The authors’ tests show that it is not as robust against removal-type transformations as it is against minor gain-type transformations (such as format/bitrate changes). Cropping falls under the category of “removal,” so PDQ does not have an advantage here.
TMK+PDQF (video) degrades especially visibly under cropping: evaluations show that if more than about 20% of the frame is cropped out, reliable matching becomes difficult. It is even suggested that stable hits should not be expected once cropping exceeds 10%, highlighting its sensitivity to substantial loss of field of view.
Comparison of Watermark Resistance
pHash is sensitive to large opaque watermarks because watermarks alter the global low-frequency structure. Traditional global DCT perceptual hashing usually struggles to maintain approximate hash consistency under strong watermarks.
PDQ has some tolerance for “light watermarks/small logos,” but under “strong opaque watermarks with complex edges and color changes,” more than half of samples may be missed at commonly used thresholds, showing limited robustness against heavy watermarks.
In video scenarios, TMK+PDQF has also been reported to perform poorly on transformations involving both watermarks and cropping. Compared with transformations that preserve the global structure, such as bitrate, format, and scaling changes, watermark overlays significantly reduce matching confidence.
Practical Notes
Platform-side experience shows that video transcoding transformations such as “format/bitrate/scaling” are handled relatively robustly, but all three algorithms show obvious performance degradation under heavy cropping or strong watermarks, especially opaque watermarks covering large areas. They should therefore be combined with upstream and downstream strategies, such as threshold tuning and manual review.
If robustness against cropping must be improved, “block-based/local” approaches can be combined, or spatial encoding methods designed for cropping/rotation can be used to compensate for the shortcomings of traditional global hashing. These can then be fused multimodally with fingerprints such as PDQ/TMK to improve recall. Both practice and research indicate that traditional global perceptual hashing is not robust against cropping, which has led to new methods designed specifically for crop robustness.
For watermark scenarios, official PDQ materials state that it can tolerate “light watermarks/small logos,” while strong opaque watermarks significantly affect matching. TMK+PDQF shows similar behavior on the video side, so relying on a single threshold alone should be avoided. When necessary, candidate expansion and a two-stage verification process should be introduced.
Key Conclusion
For the two types of local modifications, “cropping/local occlusion” and “watermark overlay,” the overall pattern among the three is as follows: PDQ and TMK+PDQF are relatively more robust against lightweight watermarks such as corner marks/logos, but both degrade significantly under obvious cropping and large opaque watermarks; pHash, because it relies more on global low-frequency features, is the least robust against these local modifications.
Friends who repost videos, now you know what to do.