hitzhangyu / self-supervised-image-enhancement-network-training-with-low-light-images-only Goto Github PK

你好，请问您知道文章中用到的图像评测指标LOE的具体代码吗

Line 71 in main.py (copied below) computes max channel for low contrast images using the ground-truth images (high_im)!! But it is supposed to be self-supervision? Am I missing something?

train_low_data_max_chan = np.max(high_im,axis=2,keepdims=True)

直接下载代码，测试出来的增强图像是灰暗的

您好，打扰您了，想问一下你这个环境是TF的那个版本，麻烦您注明下，谢谢

使用给出的代码跑了以后，出现了连接形状错误， ConcatOp : Dimensions of inputs should match: shape[0] = [1,366,490,64] vs. shape[1] = [1,365,490,64]

为什么我用jpg格式的图像进行测试，产生的图像打不开

is the hiseq applied to gray image or color image?

在model.py的train函数中，对训练数据进行shuffle时只打乱了输入数据的顺序，没有对train_high_data和train_low_data_eq打乱顺序，是我理解的有问题吗？

Hi, I have two questions regarding the implementation details.

1. Fig. 2. shows the R and I components decomposed by your network. But how do you output an enhanced image? Do you use another sub-network to enhance the illumination map?

2. For the histogram equalization loss, can you give some ideas on how to implement that?

Thank you.

Does not work with tensorflow version 2.3

您好，我是一个深度学习小白，在用您训练好的模型直接测试时，遇到的下面的错误：
tensorflow.python.framework.errors_impl.InvalidArgumentError: ConcatOp : Dimensions of inputs should match: shape[0] = [1,516,884,64] vs. shape[1] = [1,515,883,64] [[node DecomNet/concat…………
InvalidArgumentError (see above for traceback): ConcatOp : Dimensions of inputs should match: shape[0] = [1,516,884,64] vs. shape[1] = [1,515,883,64]
请问这是什么原因呢？我用的是tf1.13，py3.7。

if convenient,share your code please.thx

Since the main idea of your paper is named self-supervised image ,i am not sure what is the role the train_high_data plays in the code.Could you help me to understand?

Hello, I am very interested in your paper Self-supervised Image Enhancement Network: Training with Low Light Images Only. It is a very good work! I wonder how to implement it in code. So do you guys have any plan to public your code?

这个提亮的程度可以控制吗，感觉有点过度提亮导致噪声较多。是不是调整哪一个loss函数降低？

hi yu,
I tried your code and find some minor issues.

1. pylab module in utils.py is not given. But it seems not necessary and can be commented.
2. The channel in Decom.conv0 should be represented as 'channel//2' to keep this variable an interger. The usage of python3.7 may result in this.

一个发散的问题，这篇文章是不是重在自监督的构建而不是网络架构的选择，例如unet结构效果会不会更好？感觉你的网络结构也可以换一下对吧？