Official implementation code for "Boosting Visual Recognition in Real-world Degradations via Unsupervised Feature Enhancement Module with Deep Channel Prior"

Deep Channel Prior

Deep Channel Prior (DCP) illustrates that the channel correlation matrix of features is an explicit means to reflect the corruption type of degraded images, while the feature itself can not represent its degradation type. DCP provides an explicit optimization direction for the unsupervised solution space by reducing the difference between the channel correlation matrices of degraded features and clear features.

Prepare data and weights

You can download the training data (100 unpaired images) and test data (50,000 foggy images) of ImageNet-C (fog5) from the following path, as well as the corresponding weights： https://drive.google.com/drive/folders/1Q84HLFpjAFq91NG21piR-doxCTQHPU3q?usp=drive_link

Structure of generators and discriminators

For the generator, we employed two different structures overall. Specifically, in terms of the experiments on real datasets, we mainly utilized a U-Net style generator structure. The encoder consists of the input convolutional layer, two downsampling layers, and the dilated convolutional layer, while the decoder comprises two upsampling layers and the output convolutional layer. In terms of the experiments on synthetic datasets, we mainly adopted a flattened generator structure with residual connections. It consists of two basic blocks and a 1×1 convolutional layer, with each basic block containing several convolutional layers. For the discriminator, we exclusively utilized a PatchGAN discriminator, which comprises three convolutional layers and a sigmoid output layer.

Training

python train_VGG16.py

python train_Resnet50.py

Testing

python test_VGG16.py

python test_Resnet50.py