- 期刊
Research on Unsupervised Coloring Method of Chinese Painting Based on an Improved Generative Adversarial Network
摘要
Chinese painting elements have excellent application prospects in animation. In order to improve the efficiency and effect of image coloring, an image coloring algorithm for enhancing the generative adversarial network studied. In the original model Pix2PixHD (High‐Resolution Image Synthesis and Semantic Manipulation with Conditional GANs) first introduced a self‐attention mechanism in the generator to improve the resolution of the image, and then used VGG loss and loss score from critic as the loss function of the generator to enrich the color diversity of the image, and then continue to train the network. The size of the input image is changed, and the learning rate is adjusted to improve the training efficiency and generalization effect. Finally, the network structure optimized by using the local corresponding regularization Pixel‐Normalization. Experiments show that the algorithm can simultaneously color the high‐resolution image and preserve the underlying contour information of the original image. The subjective visual evaluation and objective quantization index (MSE, PSNR, and SSIM) are better than Pix2Pix. Image‐to‐image translation with conditional adversarial networks), Pix2PixHD and CycleGAN (Cycle‐Consistent Adversarial Networks) methods.