Analysis of Emoji Generation Based on DCGAN Model
Chenrui Xu · 2021
Deep Convolution Generative Adversarial Networks consist of a generative neural network and a discriminative neural network. The generator's objective is to learn the features of existing images and output fake samples that imitate the training data. At the same time, the discriminator attempts to discriminate the generated images from the real training data. However, the training process of the model is unstable due to the simultaneous training of two adversarial networks: the modification of one network means changing the entire model. To further understand this issue, this paper constructed a DCGAN model, which can be utilized to generate fake cartoon emojis. This paper proposed a generative neural network which generates image samples from a random input. The output of the generator and the training data, in this case the emoji images, are then given to a discriminative neural network which classifies its input as fake or real. The loss of the discriminator is then calculated based on its precision, while the loss of the generator is calculated based on the output of the discriminator. If the discriminator correctly recognizes the generated images as fake, then the loss of the generator is high; otherwise, the loss of the generator is low. The process of backpropagation updates the parameters in the discriminator and the generator using gradient descent, improving their outputs and minimizing their loss functions. As for the results, the model successfully learned the crucial features of the training data and converted the random noise to emoji images.