An experimental PyTorch implementation of DB-CNN is released at https://github.com/zwx8981/DBCNN-PyTorch! Only support experiment on LIVE IQA right now, other datasets will be added soon!
Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang
IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), to appear, 2019.
https://ece.uwaterloo.ca/~k29ma/papers/19_TCSVT_DB-CNN.pdf
Usuage:
distorted_img = distortion_generator( img, dist_type, level, seed )
Where img is the original pristine image, dist_type refers to a specified distortion type ranging in 1~9.
1, Gaussian Blur
2, White Noise
3, JPEG Compression
4, JPEG2000 Compression
5, Contrast Change
6, Pink Noise
7, Image Color Quantization with Dither
8, Over-Exposure
9, Under-Exposure
level is a specified degradation level range in 1~5.
seed should be fixed to be 1.
Running the run_exp.m script to train and test on a specifid dataset across 10 random splits.
Prerequisite: Matlab(We use 2017a), MatConvNet (We use 1.0-beta25)， vlfeat(We use 0.9.2)
Pretrained s-cnn model is included in dbcnn\data\models, you should download vgg-16 model from http://www.vlfeat.org/matconvnet/pretrained/ and put it in dbcnn\data\models.
Relevant links:
Waterloo Exploration Database: https://ece.uwaterloo.ca/~k29ma/exploration/
PASCAL VOC 2012: http://host.robots.ox.ac.uk/pascal/VOC/voc2012/