DCL-CGAN: a co-modulation generative adversarial network combined with the dual contrastive loss function for image de-occlusion
Yongjun Zhang, Yu Zhang · 2023
With the advancement of deep learning, image restoration technology has also made great progress. Most of the commonly used networks adopt encoder-decoder architecture, in which the long-term correlation between pixels can only be expressed by a deep convolutional layer. However, this method cannot guarantee the stability of the network, because the multi-level cooperative work of the network is the basic framework of the GAN network; the extremum-extreme formula and the gradient of alternating up and down make this unstable phenomenon more serious. In order to solve this problem, a co-modulation mechanism that supports remote modeling across image regions is added to the GAN model. At the same time, the traditional logistic loss function leads to insufficient generalization of the feature representation of the discriminator, which cannot stimulate the adversarial evolution of the generator, and it is easy to forget previous data pattern and the previous task . In response to this problem, this project will use the dual contrastive loss mechanism to avoid problems such as damage to the semantic structure of the generated samples or failure of the generated distribution mode. By combining these methods, the performance of PSNR, SSIM, and FID indicators for image restoration has been improved.