Awesome

Which Has Better Visual Quality: The Clear Blue Sky or a Blurry Animal?

Description

SFA code for the following papers:

• Li, Dingquan, Tingting Jiang, Weisi Lin, and Ming Jiang. "Which Has Better Visual Quality: The Clear Blue Sky or a Blurry Animal?." IEEE Transactions on Multimedia, vol. 21, no. 5, pp. 1221-1234, May 2019.
• Li, Dingquan, Tingting Jiang, and Ming Jiang. "Exploiting High-Level Semantics for No-Reference Image Quality Assessment of Realistic Blur Images." Proceedings of the 2017 ACM on Multimedia Conference. ACM, 2017.

Requirement

Framework: Caffe 1.0 + MATLAB 2016b Interface

The PLSR model uesd in the test code is trained on LIVE gblur images with DMOS (the larger the worse). w and best_layer in the journal extension are determined by five cross-validation (See TMMinter.m).

The ResNet-50-model.caffemodel is downloaded from KaimingHe/deep-residual-networks and it should be pasted into the directory models/ before you run the code! It's about 100MB which is too large to upload to this repo. If you have difficulty, you can also download the ResNet-50-model.caffemodel in my sharing on BaiduNetDisk with password u8sd.

New! We provide the PyTorch implementation of the method in SFA-pytorch

Notes

Note for training

All we need to train is a PLSR model, where the training function is plsregress in MATLAB. The features are extracted from the DCNN models pre-trained on the image classification task.

Update: remember to change the value of "im_dir" and "im_lists" in data info.

Note for datasets

You can download the datasets used in the papers from their owners for research purpose. If you have difficulty, you can refer to my sharing on BaiduNetDisk with password cu9j. We only consider the blur related images in this work.

Note for cross dataset evaluation

The reported Spearman correlation (SROCC) is multiplied by -1 when the training and testing datasets have different forms of subjective scores, i.e., one is MOS and the other is DMOS. This is to make sure that the prediction monotonicity is better when SROCC is closer to 1.

Citation

Please cite our papers if it helps your research:

[Paper]

[Paper] [Poster]

Contact

Dingquan Li, dingquanli AT pku DOT edu DOT cn.