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!
Blind Image Quality Assessment Using A Deep Bilinear Convolutional Neural Network (Official IEEE preprint version)
Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang https://ieeexplore.ieee.org/document/8576582
IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), to appear, 2019.
Deep Bilinear Pooling for Blind Image Quality Assessment (Unofficial free version)
Files under distorion_generator are used for synthesizing distorted images.
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
level is a specified degradation level range in 1~5.
seed should be fixed to be 1.
Training codes live in dbcnn folder.
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.