the brisque model seems yield wrong result quite often

krshrimali / No-Reference-Image-Quality-Assessment-using-BRISQUE-Model

Implementation of the paper "No Reference Image Quality Assessment in the Spatial Domain" by A Mittal et al. in OpenCV (using both C++ and Python)

MIT License

192 stars 35 forks source link

the brisque model seems yield wrong result quite often #6

Open braindevices opened 2 years ago

braindevices commented 2 years ago

the brisque model seems yield wrong result quite often. I have original png image, jpeg compressed with 90% quality, jpeg with 60% quality. The computed score is:

(23.201696395874023, 0.0, 0.0, 0.0) for png
(19.102746963500977, 0.0, 0.0, 0.0) for high quality jpg
(17.447845458984375, 0.0, 0.0, 0.0) for low quality jpg

If I get the paper correct, the higher the score the lower the quality.

png: good_small

high quality jpeg: good_small_high

low quality jpeg: good_small_low

braindevices commented 2 years ago

is there any explanation on this behaviour? Is there any assumption about the image content does not meet resulting in this failure?

braindevices commented 2 years ago

I actually used the model and object from opencv contrib.

krshrimali commented 2 years ago

Hi, @braindevices - I'm sorry that I couldn't respond to this. Just to confirm, can you try using the OpenCV's quality module once? And share the scores you get? https://docs.opencv.org/4.x/d8/d99/classcv_1_1quality_1_1QualityBRISQUE.html