Replies: 2 comments 5 replies
-
|
The biggest problem is that the the hash algorithm is incredibly easy to change, if you resize before you turn to grayscale is different than turning grayscale before you resize. For example webp files there is a simple option that changes how it converts yuv/cbcr to RGB which then screws up the exact values when it resizes and converts to grayscale. Even if ImageHash doesn't change, to keep the hash the same it depends on the version of PIL for resize and grayscale conversion and the version of libwebp or libjpeg or whichever library loads the image and if the library has a direct conversion to RGB or if PIL has to do it itself (PIL always converts images into RGB(A) images) |
Beta Was this translation helpful? Give feedback.
-
|
Sounds like you've done some testing? How much of a difference is there in those different scenarios? I thought one purpose of the process was to make things "close enough" for matching? Is it significant enough to effect the hamming distance enough? I would presume that even with the variables, the distance is not enough to make a difference? We don't need the exact same number, just one close enough for confidence? |
Beta Was this translation helpful? Give feedback.
-
|
I don't have any conclusive testing but the Ahash which is the most prone to false-positives. In otherwords more images will have the same Ahash than Dhash or Phash. On my test of a single webp image with differences in resize/grayscale order a slightly different resize algorithm and incorrect color conversion it could not produce the same Ahash and the Dhash and Phash showed even more differences. After correcting the resize/grayscale order the resize algorithm and getting a little bit more accurate color conversion the Ahash matches and the Dhash hamming distance was ~5. Would they have matched? Yes. But this is the same exact file sent through both hash algorithm. When using a different file with actual differences it may be even less accurate. I plan on doing some better testing in the future but I want at least get webp to work exactly for my setup before I do that. |
Beta Was this translation helpful? Give feedback.
-
|
In this case I think we would just use the hash in place of the cover image, however CT currently does not have a phash implementation and I'm not sure I want to pull in numpy and scipy as further binary dependencies right now. I recently noticed that after some of my changes since I've taken over as maintainer for CT have drastically increased initial load time for at least the CLI and I want to see what I can do to fix that |
Beta Was this translation helpful? Give feedback.
-
|
I did find this pure python for DCT https://gist.github.com/vilcans/8531160 and from what I remember from comparing the aHash the numpy isn't needed. |
Beta Was this translation helpful? Give feedback.
-
|
A hash and dhash are incredibly simple and don't need numpy in the slightest though it probably makes them a tad faster. If you are able to test it that would be great. There is some code for a perceptual algorithm that is commented out in imagehash.py I'm unsure if it is similar to the ImageHash library |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Metron has now added pHash for its covers. My current thinking is:
hashto ComicIssue (knowing the hash type will be required so maybe a JSON field?{"type": "phash", "hash": "hexhash"}Possibly support multiple hash types? Another option could be a string (array):p_hexhashora_hexhashetc.)I'm wondering if the hash from the API should be ignored? If the list of possible issues does not match, maybe download the image and hash it ourselves to double check?
I'm also wondering if saving the hash of the local (to issue) or remote (to image cache) would be of any use?
When I comes to the hashing, should it be switched to the imagehash or implement the same in CTs? I notice that it requires a few deps that for CTs use case aren't really required?
Beta Was this translation helpful? Give feedback.
All reactions