zwx8981 / dbcnn

Blind Image Quality Assessment Using A Deep Bilinear Convolutional Neural Network

Weixia Zhang (张维夏)👋

I am an Associate Research Scientist at AI Institute, Shanghai Jiao Tong University. Currently, I work on perceptual quality evaluation and enhancement for visual content produced by various manners, i.e., PGC, UGC, and AIGC.

dbcnn's Issues

训练时出现DEROUTPUT dimensions are incompatible with X and FILTERS

大神您好！同前几个同学一样，我在运行run_exp.m的时候也出现这个问题。
错误使用 vl_nnconv
DEROUTPUT dimensions are incompatible with X and FILTERS.

出错 vl_simplenn (line 413)
vl_nnconv(res(i).x, l.weights{1}, l.weights{2}, res(i+1).dzdx, ...

出错 cnn_train>processEpoch (line 350)
res = vl_simplenn(net, im, dzdy, res, ...

出错 cnn_train (line 150)
[net, state] = processEpoch(net, state, params, 'train') ;

出错 initializeNetworkTwoStreams (line 280)
[netc, info] = cnn_train(netc, bcnndb, @getBatch_bcnn_fromdisk, opts.inittrain, ...

出错 imdb_bcnn_train_dag (line 70)
net = initNetFn(imdb, encoderOpts, opts);

出错 run_experiments_bcnn_train (line 36)
imdb_bcnn_train_dag(imdb, options);

出错 run_exp (line 63)
[options, imdb] = run_experiments_bcnn_train(opts);
我看您没有回答。我把读取图像的路径改成我自己的了，matconvnet与你版本是一样的，另外我是单GPU，把那个编号由2改为1了，其它没有动。我GPU是1070ti，请大神指教！谢谢！

关于程序

大神您好！
我在跑您的代码的过程中遇到了一个错误：
调用这一句im = vl_imreadjpeg(images,'numThreads', opts(i).numThreads)时出现：
could not read image '.../fastfading/img40.bmp' because 'libjpeg: Not a JPEG file: starts with 0x42 0x4d'
在网上也没找到可靠的答案，不知道您是否遇到这问题？

the cross dataset test problem

Hello!
I think the method proposed in this article is very novel.
I want to train the model on the LIVE dataset and test on TID2013. However,they have different label.(dmos for LIVE,mos for TID2013) I am not very familiar with matlab, so can you tell me how to convert these labels

Pre-trained model

Hello!
I haved read the article, I think the method is very novel.
But I want to run 'demo.m', so may I ask you for a pre-trained model?
When I run the 'run_exp.m', I find that 'run_experiments_bcnn_train.m' is missing, How to solve the problem？
Thank you very much!

Error in vl_imreadjpeg()

大神你好
我在重新训练 model 的时候, 在 imdb_get_batch_bcnn.m line 70 im = vl_imreadjpeg(images,'numThreads', opts(i).numThreads) 发现了error, MATLAB 没有给出任何的error message，请问您知道这是为什么吗？
Dataset: LIVE release2 (bmp image)
系统：WINDOWS 10

    if fetch
        im = vl_imreadjpeg(images,'numThreads', opts(i).numThreads) ;
    else
        im = images ;
    end

特意检查过images的path修改后是正确的，但不知道为什么跑vl_imreadjpeg的时候程序throw error

Wrong use of vl_argparse

Hello author, when I trained run_exp.m according to your steps, I found that "vl_argparse" was wrong. How should I solve this problem?

no net component in the dagnet.net from demo.m file

Thanks for sharing your code with the community. I was testing it recently, but got the error in the demo.m.

Reference to non-existent field 'net'.
Error in demo (line 7)
dagnet = dagnn.DagNN.loadobj(dagnet.net) ;

I used the
net_path = 'DBCNN\dbcnn\models\LIVE_models\whole\imdb-seed-1.mat';
Is this the correct model? Thanks a lot for your support.

Best regards!

Can your provide a link to download the Live Challenge?

Hi, ^_^. I'm here again.
I can't open the link 'https://www.cs.utexas.edu/%E2%88%BCdeepti/ChallengeDB.zip'.

Can you help me?

Which model should I use to test on the waterloo dataset?

hi~thanks very much for your excellent work. I have encountered some confusion and need your help.
You mentioned that in the paper "retrain the S-CNN stream using distorted images generated
from PASCAL VOC 2012 only to ensure the independence of image content between training and testing". Which database do you use when fine-tuning the DB-CNN?
Whether the models under the folder "models" use the waterloo dataset when training S-CNN? I want to test the results on waterloo, which model should I use?
Thanks!

您好，看到您的一些关于图像质量评价的介绍，想问您个问题，就是BLIINDS-II无参考的指标，我在github上复现了几个matlab的代码，文章中介绍BLIINDS-II越大表明图像质量越好，但是我复现的代码结果是相反的，请问您有这个BLIINDS-II代码吗？

关于训练的问题

作者你好，请问一下我想用自己的数据集进行训练，也必须在正式训练时生成对应的失真图像才可以进行么

where can I found the SCNN labels

Appreicate for sharing your great work first!
But I really wonder where can I found the SCNN labels. In the paper, you said there are 125 labels, and in the matlab distortion dir, there is 8 kinds of distortion and each has 5 levels. so there is 40 lables? correct me if i am wrong

error finding a file

when I run the code I get this error

train: epoch 01: 1/ 78:Error using load
Unable to read file 'data\checkgpu\live-seed-01\nonftbcnn\bcnn_nonft_00003'. No such file or directory.

could you please help me? how is this file created?