Image Colorization : GAN and Autoencoder Models’ Performance Comparison

Two basic models, the GANand Autoencoder models, are designed to color black-and-white pictures. The purpose is to compare the color data and methods of black-and-white pictures obtained by the two models through continuous training. The advantages and disadvantages of the two models are found through comparison, such as overfitting. Judging from the advantages and disadvantages, which model is more suitable for black-and-white picture coloring can train better data. The results show that from the image, it can be seen that the valuable loss of Autoencoder is better than the training loss, which is the last. When the time data is 15 ~ 35, there is an error. In contrast, the GAN model has serious over-fitting problems at any time node, which can not summarize and color all the possible black-and-white photo data. It is concluded that from all the advantages and disadvantages, the error of the Autoencoder model is small in all aspects. Gan model has serious errors, which may lead to incomplete data and training. As shown in the figure, this point is located below the valuable loss of the GAN model, which means that all valuable losses cannot train a reasonable and color picture through this point. Similarly, many black-and-white photo data similar to this point cannot obtain the color picture after coloring through this model. This situation is called overfitting, which is also the biggest problem of the Gan model. It shows that the Autoencoder model is more suitable for black-and-white picture coloring.