VIDGIQA

May 7, 2018 ยท View on GitHub

Code of paper ``Visual Importance and Distortion Guided Deep Image Quality Assessment Framework"

Please cite as:

Jingwei GUAN, Shuai YI, Xingyu ZENG, Wai-Kuen CHAM and Xiaogang WANG. "Visual Importance and Distortion Guided Deep Image Quality Assessment Framework." IEEE Transactions on Multimedia.

  • Usage:
  1. Open Matlab
  2. unzip VIDGIQA_v1.0.zip.
  3. Run Demo.m
  • Note:
  1. The code loads a non-distortion specific model trained on the LIVE dataset. The training samples used are the training subset' illustrated in the Evaluation on the LIVE dataset' Section.
  2. The code is an unoptimized CPU version written with Matlab.
  3. The computational time of processing 512*512 RGB image using VIDGIQA_1 is about 39.041sec; while adding data augmentation in the testing procedure (using VIDGIQA), the computational time is about an hour. With the Python library Theano with Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10GHz CPU, GeForce GTX TITAN Black GPU, the computational time of the proposed method is 0.042s, while adding data augmentation in the testing procedure, the computational time is 2.325s.