About the results of fbcnn_color.pth

[YangGangZhiQi]

Hi,
Great works! I test your model(fbcnn_color.pth) in 'testset/Real' dataset. The results are not so remarkable as the picture in this master. The output of model(fbcnn_color.pth) without qf_input are as follow(left is input, right is output):

I don't kown if there are something wrong with my results. And the output of model(fbcnn_color.pth) with qf_input are also not so good. When zoom out, I can find obvious artifacts. Hope for your reply.

[jiaxi-jiang]

Hi, thanks for your interest!

The provided color model is trained with the single JPEG degradation model. As we analyze in the paper, such blind models usually can not deal well with real JPEG images which are compressed multiple times. But our FBCNN is a flexible model, to get the desirable results with fewer artifacts, just set qf_input as a smaller number, e.g. 10. See https://github.com/jiaxi-jiang/FBCNN/blob/main/main_test_fbcnn_color_real.py#L23

BTW, to get the desired result automatically, you can either estimate the dominant smaller quality factor using the method proposed in our paper (FBCNN-D), or augment the training data with our proposed double JPEG degradation model (FBCNN-A).

[jiaxi-jiang]

The pictures I show are exactly the same as 'test/Real'. Just directly run main_test_fbcnn_color_real.py then you can get the result. Please note that the quality factor [0,100] is transformed to [1,0] as the input of the network. So quality factor 10 corresponds to 0.9 as the input parameter.
https://github.com/jiaxi-jiang/FBCNN/blob/main/main_test_fbcnn_color_real.py#L87
Let me know if you still can not get the same result.

[jiaxi-jiang]

I don't know your aim of coming here. The images I show in this repo is super easy to get from the provided codes, pre-trained model and test images. Malicious comments are not welcome.

Fig.1 with input qf_control = 10

Fig. 2 with input qf_control = 5

Fig. 3 with input qf_control = 10

Fig. 4 with input qf_control =10

Fig. 5 with input qf_control = 70

Fig. 6 with input qf_control = 70

log file

Real_fbcnn_color.log

[jiaxi-jiang]

I just download my repo from a new machine and run the code, without changing anything. The models and test images are both exactly the same. The results you show with many artifacts seem that the qf_control is a large number e.g. 90. So I am interested to see your results with qf_control = 5 or 10 because in this case most artifacts together with some texture details should be removed.

[cszn]

Just run the code main_test_fbcnn_color_real.py.

python main_test_fbcnn_color_real.py

[YangGangZhiQi]

@jiaxi-jiang @cszn When running the model with cpu mode, I can get the same results with yours. But with gpu mode, the output pictures of model with qf_control = [5,10, 30, 50, 70, 90] are same. I print the values of qf_embedding in FBCNN network with qf_control=[5,10,50, 70, 90], there are same too. It's really weird that the outputs become different with different device(cpu/gpu).