A Method based on Evolutionary Algorithms and Channel Attention Mechanism to Enhance Cycle Generative Adversarial...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) enable image-to-image translation between domains.
• Cycle-consistent Generative Adversarial Networks (CycleGAN) are popular but face training instability and low-fidelity image generation issues.
• Existing CycleGAN improvements often modify network structures or loss functions, with limited success in feature discrimination.

Purpose of the Study:

• To address the training instability and low fidelity in CycleGAN models.
• To enhance the generator's ability to capture discriminating features for higher quality image synthesis.
• To introduce a novel method combining Evolutionary Algorithms and Attention Mechanisms for GAN training.

Main Methods:

• Proposed Attention Evolutionary GAN (AevoGAN) integrates Evolutionary Algorithms (EAs) with Attention Mechanisms.
• EAs progressively optimize generator weight activation vectors from an initial CycleGAN.
• Channel attention mechanism is employed to help generators learn crucial image features.

Main Results:

• AevoGAN alleviates training instability issues inherent in CycleGAN.
• The method generates higher quality images with improved texture details compared to standard CycleGAN.
• Quantitative evaluation shows superior performance over existing CycleGAN methods using Inception Score (IS), Fréchet Inception Distance (FID), and Kernel Inception Distance (KID).

Conclusions:

• AevoGAN offers a robust solution for unpaired image-to-image translation.
• The combination of EAs and attention mechanisms significantly improves GAN training stability and output image quality.
• This approach advances the state-of-the-art in GAN-based image synthesis.