A Novel Generator With Auxiliary Branch for Improving GAN Performance

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) typically generate images using a coarse-to-fine approach, with generators refining details through multiple layers.
• Existing GAN generator architectures, often relying on stacked residual blocks, can hinder information flow, impacting image generation quality.
• Efficient propagation of coarse image information to later layers is crucial for high-fidelity image synthesis in GANs.

Purpose of the Study:

• To introduce a novel GAN generator architecture designed to overcome information flow limitations in traditional residual block-based generators.
• To enhance the coarse-to-fine image generation process by effectively combining features from different network branches.
• To improve the stability and quality of GAN-generated images through a more efficient information propagation mechanism.

Main Methods:

• Proposed a novel generator architecture featuring two distinct branches: a main branch utilizing residual blocks and an auxiliary branch for conveying coarse information.
• Introduced a gated feature fusion module (GFFM) to intelligently control and combine feature information from both the main and auxiliary branches.
• Conducted extensive experiments on diverse datasets (CIFAR-10, CIFAR-100, LSUN, CelebA-HQ, AFHQ, tiny-ImageNet) and performed ablation studies to validate the method's effectiveness and generalization.

Main Results:

• The proposed GAN generator architecture demonstrated superior performance compared to existing methods on multiple standard image generation datasets.
• Quantitative evaluations showed significant improvements in image quality metrics, including Inception Score (IS) and Frechet Inception Distance (FID).
• Specifically, on the tiny-ImageNet dataset, the FID score improved from 35.13 to 25.00, and the IS score increased from 20.23 to 25.57.

Conclusions:

• The novel generator architecture effectively addresses the information flow problem in GANs, leading to enhanced image generation capabilities.
• The gated feature fusion module plays a critical role in successfully integrating features from parallel branches, optimizing the generation process.
• The proposed method offers a promising advancement in GANs, achieving state-of-the-art results and demonstrating strong generalization across various datasets.