Visual Performance of DCGAN Model for Analyzing Emoji Generation
Wenyan Xu, Zhenheng Xu · 2022 IEEE 2nd International Conference on Data Science and Computer Application (ICDSCA) · 2022
As social media is rapidly developing among the young generations, emojis are becoming increasingly popular in recent years, establishing the need to expand emojis. However, continuously creating new emojis is hard to accomplish due to potential excessive costs and time-consuming. This paper introduces a DCGAN model to tackle this issue, which can generate new emojis by simply learning from the existing emojis input into the model. This application is built up mainly using the Pytorch library in Python. The implementation uses the generator to create plausible emojis to cheat the discriminator by training the model with input real emojis. Then with the results of their operation, these two models are trained repeatedly to improve their abilities. Thus, with this adversarial operation, the generator can produce emojis that have the reasonably authentic quality to pass the test from the discriminator, while the discriminator's accuracy is also continually improving during the training. Eventually, a GUI is created as an image gallery to display the implemented emojis using Tkinter in Python. Overall, this DCGAN model is entirely satisfactory to create new emojis randomly with good quality. However, those emojis generated by this implementation have some quality defects that still need to be refined manually, such as disordered or unclear face features. Therefore, this model can be used as a basic emoji generation program simply for entertainment or commercial use, while emojis inevitably require editing by designers.