[Image Processing] Looking for non-reference based image quality check, BRISQUE results seems untrustworthy on small patch.

Hi,

I'm having a noisy volumetric 3d data[cubical] of a coin with dark background

like 360x360x2048 = image [360x360] = max(data,[],3)

Algo1: get denoisy data A [takes 1hour to run]

Algo2: get denoisy data B [takes 1hour to run]

Algo3: get denoisy data C [takes 1hour to run]

data A looks better compared to rest and brisque value also says the same [brisque ranging from 15-30]

to work faster we tried to work on patch of 60x60, 120x120, 240x240, 256x256 time ranging from [4 mins to 30mins]

Algo1: get denoisy data A

Algo2: get denoisy data B

Algo3: get denoisy data C

data A looks better compared to rest but now brisque value seems to give random numbers [40-120] and not even in ordered way.

We are looking for a non-reference based image comparison thing.

Please help what all we should explore?

Answer 1

[Image Analyst]

Try niqe and piqe.

And see https://en.wikipedia.org/wiki/Image_quality

And see articles in the Image Processing Literature on it:

• 5.3.10 Image Quality Evaluation, Visual Quality, Quality Assessment, and Imaging Models
• 5.3.10.1 Learning for Image Quality Evaluation, CNN, GAN
• 5.3.10.2 No-Reference Image Quality Evaluation
• 5.3.10.3 Color Image Quality, Hyperspectral Image Quality
• 5.3.10.4 Screen Content Image Quality Evaluation
• 5.3.10.5 Image Quality Evaluation, Stereoscopic Imagery, Stereo, Depth
• 5.3.10.6 Image Quality Evaluation, Perceptual Quality, Subjective Quality, Human Visual System Based, HVS
• 5.3.10.7 Compressed Image Quality Evaluation, JPEG Quality Evaluation
• 5.3.10.8 Image Quality Evaluation, Geometric Quality, Spatial Distortions
• 5.3.10.9 Noise Models, Digitization Noise
• 5.3.10.9.1 Noise Models, Noise Level
• 5.3.10.10 Video Quality, Video Image Quality Evaluations
• 5.3.10.11 Video Quality for Stereo, 3D Video
• 5.3.10.12 Video Quality for Compression, Coded Video