Home

Awesome

Learning Semantic Degradation-Aware Guidance for Recognition-Driven Unsupervised Low-Light Image Enhancement (AAAI2023)

Paper link

How to train SDAG

python train.py -opt /SDAG/options/train/SemanticAwareRecon.yml

How to integrate the SDAG into an existing enhancer

  1. Define the SDAG network and load the pre-trained model.
SemanticNet = SemanticAwareNet(channels=64).cuda()
SemanticNet.load_state_dict(torch.load('/gdata1/zhengns/checkpoint/PersonalizedEnhancement/experiments/SemanticAware_margin05_lqRecon_NetC_selfSupervised/models/net_g_78000.pth')['params'])
  1. Self-reconstructing the enhanced image from the unsupervised enhancer for measuring degradations in the semantic level.
semanticPrior = SemanticNet(enhanced_image) #
  1. Employing SDAG as a regularization term.
loss_semantic = F.l1_loss(enhanced_image, semanticPrior, reduction='mean')