C-SupConGAN: Using Contrastive Learning and Trained Data Features for Audio-to-Image Generation

In this paper, the audio-to-image generation problem is investigated, where appropriate images are generated from the audio input. A previous study, Cross-Modal Contrastive Representation Learning (CMCRL), trained using both audios and images to extract useful audio features for audio-to-image generation. The CMCRL upgraded the Generative Adversarial Networks (GAN) to achieve high performance in the generation learning phase, but the GAN showed training instability. In this paper, the C-SupConGAN that uses the conditional supervised contrastive loss (C-SupCon loss) is proposed. C-SupConGAN enhances the conditional contrastive loss (2C loss) of the Contrastive GAN (ContraGAN) that considers data-to-data relationships and data-to-class relationships in the discriminator. The audio and image embeddings extracted from the encoder pre-trained using CMCRL is used to further extend the C-SupCon loss. The extended C-SupCon loss additionally considers relations information between data embedding and the corresponding audio embedding (data-to-source relationships) or between data embedding and the corresponding image embedding (data-to-target relationships). Extensive experiments show that the proposed method improved performance, generates higher quality images for audio-to-image generation than previous research, and effectively alleviates the training collapse of GAN.