UGAN-2G：An Unsupervised Target Image Remodeling Enhancement Network

In facing of the generative adversarial network (GAN) has slow training convergence speed, fuzzy generated image features, and network information separation leads to the lack of image diversity. Combined with U-Net network architecture, this paper proposed a new target image remodeling enhancement algorithm for a dual-gradient U-type generative adversarial network( Encoder10 decoder Generative Adversarial Network with Gradient Fusion and Gradient Penalty, UGAN-2G), which was an unsupervised self-learning generative model, to solve the problem of target image detail loss while enhancing the multitasking of target detection. This method chooses a WGAN-GP network with a gradient penalty, which solves the instability of the generative adversarial network. Moreover, a new generator, the U-type network with a self-mapping gradient fusion, was proposed to lock the target image, which the fusion of the DenseNet idea to break the symmetry of the network, the unsupervised remodeling of the details of the finished image, and improve the network usage efficiency. To enhance the remodeling effect, combine the image pixel loss with the adversarial loss as the generator loss, and the trained model is more robust. Comparison of the algorithm performance in the FFHQ-baby, LSUN, and pavement crack datasets. The FID (Frechet inception distance) values of UGAN-2G decreased by 29.5% and 25.8% over conventional WGAN-GP and DCGAN algorithms, and decreased by 16.3% and 10.1% over deep convolutional UNet-GAN and Re-GAN algorithms, indicating better image quality from UGAN-2G. The IS (Inception score) values of UGAN-2G increased by 23.3% and 21.6% over WGAN-GP and DCGAN algorithms, respectively, and increased by 20% and 14.1% over UNet-GAN and Re-GAN algorithms, indicating a good diversity of images generated by UGAN-2G. The experimental results show that UGAN-2G algorithm integrates deep convolution to enhance learning ability, with rich application scenarios, high target image quality and strong diversity, and accelerates network convergence with good robustness.