School of Software, Nanjing University of Information Science and Technology, Nanjing 210044, P. R. China.
NICE Research Group, Department of Computer Science, University of Surrey, Stag Hill Campus, Guildford GU2 7XH, UK.
Int J Neural Syst. 2023 May;33(5):2350026. doi: 10.1142/S0129065723500260. Epub 2023 Apr 5.
A Generative Adversarial Network (GAN) can learn the relationship between two image domains and achieve unpaired image-to-image translation. One of the breakthroughs was Cycle-consistent Generative Adversarial Networks (CycleGAN), which is a popular method to transfer the content representations from the source domain to the target domain. Existing studies have gradually improved the performance of CycleGAN models by modifying the network structure or loss function of CycleGAN. However, these methods tend to suffer from training instability and the generators lack the ability to acquire the most discriminating features between the source and target domains, thus making the generated images of low fidelity and few texture details. To overcome these issues, this paper proposes a new method that combines Evolutionary Algorithms (EAs) and Attention Mechanisms to train GANs. Specifically, from an initial CycleGAN, binary vectors indicating the activation of the weights of the generators are progressively improved upon by means of an EA. At the end of this process, the best-performing configurations of generators can be retained for image generation. In addition, to address the issues of low fidelity and lack of texture details on generated images, we make use of the channel attention mechanism. The latter component allows the candidate generators to learn important features of real images and thus generate images with higher quality. The experiments demonstrate qualitatively and quantitatively that the proposed method, namely, Attention evolutionary GAN (AevoGAN) alleviates the training instability problems of CycleGAN training. In the test results, the proposed method can generate higher quality images and obtain better results than the CycleGAN training methods present in the literature, in terms of Inception Score (IS), Fréchet Inception Distance (FID) and Kernel Inception Distance (KID).
生成对抗网络 (GAN) 可以学习两个图像域之间的关系，并实现非配对的图像到图像的翻译。其中一个突破是循环一致性生成对抗网络 (CycleGAN)，这是一种将源域的内容表示转换到目标域的流行方法。现有研究通过修改 CycleGAN 的网络结构或损失函数，逐渐提高了 CycleGAN 模型的性能。然而，这些方法往往存在训练不稳定的问题，并且生成器缺乏获取源域和目标域之间最具区分性特征的能力，从而导致生成的图像保真度低且纹理细节少。为了克服这些问题，本文提出了一种结合进化算法 (EAs) 和注意力机制来训练 GAN 的新方法。具体来说，从初始的 CycleGAN 开始，通过 EA 逐步改进表示生成器权重激活的二进制向量。在这个过程的最后，可以保留生成器表现最好的配置用于图像生成。此外，为了解决生成图像的保真度低和缺乏纹理细节的问题，我们使用了通道注意力机制。后者使候选生成器能够学习真实图像的重要特征，从而生成质量更高的图像。实验从定性和定量两个方面证明了，所提出的方法，即注意力进化 GAN（AevoGAN）缓解了 CycleGAN 训练的训练不稳定问题。在测试结果中，所提出的方法可以生成更高质量的图像，并在 Inception Score（IS）、Fréchet Inception Distance（FID）和 Kernel Inception Distance（KID）等方面获得比文献中现有的 CycleGAN 训练方法更好的结果。
