Multi-class GAN for generating multi-class images in object recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative Adversarial Networks (GANs) face limitations in data augmentation for object recognition, including unstable training and poor image quality.
• Existing GAN variants like progressive growing GANs, multi-scale gradient GANs, and packed GANs (PacGAN) address specific issues but cannot generate multiple image types simultaneously and have long training times.

Purpose of the Study:

• To address the limitations of current GANs in multi-class data augmentation for object recognition.
• To propose a novel Multi-class Generative Adversarial Network (Mc-GAN) that overcomes the constraints of single-image generation and lengthy training.

Main Methods:

• Introduction of the Multi-class Generative Adversarial Network (Mc-GAN).
• Utilizing an augmented discriminator to concurrently train multiple generators.
• Employing iterative training to enable the discriminator to guide each generator for accurate image synthesis.
• Analysis of the Mc-GAN objective function's optimization process.

Main Results:

• Mc-GAN successfully generates high-quality images across multiple classes.
• The proposed method significantly reduces GAN training time.
• Experimental validation demonstrates the effectiveness of Mc-GAN for object recognition data augmentation.

Conclusions:

• Mc-GAN offers a practical solution for multi-class data augmentation in object recognition.
• The approach enhances the overall practicality and efficiency of Generative Adversarial Networks.
• Mc-GAN represents a significant advancement in GAN applications for computer vision tasks.