Improve blurry issue type

[jwmueller]

Some images currently obtain undesirably low blurry quality scores, when they do not appear blurry.
In particular:

• uniformly-colored images
• high-resolution images
• [] odd aspect ratio images

Important links

• https://github.com/WillBrennan/BlurDetection2

[jwmueller]

some more images flagged as blurry (source):

[saskra]

I can confirm that blurred image detection gives way too many false positives, while the actual blurred ones are not even among the top ranks. The other issue types work much better.

[jwmueller]

@saskra Thanks for the info! What sized images did you run on? We suspect some image sizes may not work well with our current blur check. It'd be helpful if you can share the results of "false positives" and "actual blurred ones" not detected!

• Any other ideas you have why this blurry issue type is not working optimally?

[saskra]

This is a dataset of >7,500 smartphone photos of various resolutions, of which >500 were detected as blurry, even though it should be a low double-digit number. Unfortunately, I cannot publish the images. I also can't tell yet if the actual blurred images were not detected at all, I've only checked so far that they are not in the first hits.

In fact, I haven't found a working alternative yet either, at best I can automatically detect the blurrier one between two very similar images.

[jwmueller]

Interesting, well thanks for sharing the info you can! Our team is looking into improvements for blurry-detection, but it may be challenging to do via simple algorithms for every possible image domain.

If you ever try it on other image datasets you have, do let us know how that worked!

[sanjanag]

huggingface dataset Matthijs/snacks

[sanjanag]

huggingface dataset : cats_vs_dogs

[smttsp]

@jwmueller

In my previous company, I used this one and it worked perfectly:

https://stackoverflow.com/a/20198278

---

As far as I remember, I tried a bunch of methods but none of them was as good as this

[sanjanag]

Hi @smttsp! Thanks for the suggestion! I have tried using Canny as an edge detector and further processing its results. It works great, but it is too slow for us. We had to do a tradeoff between speed and the quality of results. The current approach gives good results and runs much faster compared to the Canny detector.

[smttsp]

would you mind sharing your code if you still have it? canny edge detector should not be slow, it is just matrix multiplication. I would be surprised if the code I shared takes more than a few ms.

[sanjanag]

I'm sorry I don't have it saved anywhere, could you post some results using your strategy, both from quality and speed perspective. You can start with our example dataset or any of the datasets here

[sanjanag]

Odd aspect ratio images flagged as blurry