Twin Auxiliary Classifiers GAN.

Area of Science:

• Artificial Intelligence
• Computer Vision

Background:

• Conditional generative models have advanced significantly.
• Auxiliary Classifier GAN (AC-GAN) uses an auxiliary classifier for discriminative image generation.
• AC-GAN faces limitations in sample diversity with increasing class numbers.

Purpose of the Study:

• To theoretically identify the cause of decreased diversity in AC-GAN.
• To propose a novel method, Twin Auxiliary Classifiers GAN (TAC-GAN), to improve diversity.
• To enhance the performance of conditional image generation on large-scale datasets.

Main Methods:

• Theoretical analysis of the auxiliary classifier's impact on AC-GAN.
• Introduction of TAC-GAN with an additional interacting component.
• Minimizing divergence between generated and real data distributions.

Main Results:

• TAC-GAN effectively addresses the perfect separability issue in AC-GAN.
• The proposed TAC-GAN minimizes divergence between real and generated data distributions.
• Experimental results demonstrate TAC-GAN's ability to replicate data distributions and improve diversity.

Conclusions:

• TAC-GAN overcomes the diversity limitations of AC-GAN for large-scale conditional image generation.
• The novel architecture improves the quality and diversity of generated images.
• TAC-GAN shows significant improvements on both simulated and real-world datasets.