Article Preview
Top1. Introduction
AI-Generated Content (AIGC) (Liu, 2023; Kang, 2023) has made significant progress in content generation. Designers can now use AIGC to create emotionally captivating designs that resonate deeply with people. This paper utilizes deep learning technology to enhance the selection of emotional design through the separation of appearance and color. Deep learning technology can solve many problems, such as object detection, image classification, image segmentation (Ranjbarzadeh & Tirkolaee, 2023), and image generation, and has been widely applied in various industries (Weber, Arabnia, Aydın, Tirkolaee, 2023; Song, Li, et al., 2022; Wu, 2022; Morales & Suárez-Rocha, 2022; Dayyala, 2022). Deep learning (Miao & Ruomu, 2023; Zhou & Fang, 2022) is a machine learning subfield originating from the study of artificial neural networks. Currently, common neural networks used in deep learning include CNN, RNN, DNN, PNN, and others. These algorithms effectively model complex relationships within data through multi-layer representations. Unsupervised learning (Kumar, et al., 2021) has always posed a significant challenge for researchers. However, the rapid progress in unsupervised learning can be attributed to extensive research on generative models. These models can generate new samples based on the high-dimensional data distribution. In 2014, Ian Goodfellow proposed Generative Adversarial Networks (GANs) (Dalva, et al., 2022; Yang, et al., 2022; Kang, et al., 2023; Goodfellow, 2016; Li, et al., 2021; Zou & XiuFang, et al., 2022) which outperformed other generative models, producing better samples owing to the generation of countermeasures network. Compared to Pixel RNN (GANs) (van den Oord, et al., 2016; Salimans & Kingma, 2016), GAN generates a sample faster. In contrast to Variational Auto-Encode (VAE) (Sahu & Majumdar, 2021; Zhu, et al., 2021), a GAN does not have a lower bound of change, and its various adversarial generation networks gradually converge, while VAE exhibits some bias. Furthermore, a GAN does not possess a lower variation bound or a complex partition function like deep Boltzmann machines (Liu, et al., 2021; Luo, et al., 2021). GANs can generate samples simultaneously, eliminating the need for repeatedly applying the Markov chain solver (Karimi, et al., 2021; López-Suárez & Urriza, 2021). By comparison, it was found that GAN is more suitable for the application of this paper. Hence, this paper mainly focuses on generating content design, which requires diversification.
During the training process of a GAN, the generator network learns the distribution of real samples and determines whether an input image or real data is generated. The optimization of the entire model involves solving a “binary minimax game” problem. Currently, GANs represent the state-of-the-art in generative models. As researchers delve deeper into the realm of GANs, they have proposed several improved algorithms and innovative applications. For instance, the WGAN (Arjovsky, et al., 2017; Gulrajani, et al., 2017; Li, et al., 2021; Dong, et al., 2020; Xu, et al., 2021; Wang, et al., 2021) addresses the instability issues encountered during GAN training. DCGAN (Radford, et al., 2015; Deleforge, et al., 2021; Kim, et al., 2021) utilizes CNN in GAN to produce images of better quality. EBGAN (Chen, et al., 2016) treats the discriminator as an energy function, assigning low energy to actual data and high energy to fake data. On the other hand, InfoGAN (Liu & Lu, 2021; Gao, et al., 2021; Chen, et al., 2021) achieves decomposable feature representation through unsupervised learning. It combines a GAN with maximizing the mutual information between the generated image and the input code.