改进生成对抗网络的小样本图像数据增强算法

An Improved Small-Sample Image Data Augmentation Algorithm for Generative Adversarial Networks

摘要: 针对小样本场景下深度学习图像数据集匮乏，且现有生成式对抗网络(Generative Adversarial Networks, GAN)存在生成图像清晰度低、训练稳定性差的问题，提出一种融合相对判别器与零中心梯度惩罚机制(R₁，R₂正则化)的相对论配对生成对抗网络(Relativistic pairing GAN, RpGAN)图像数据增强算法。算法对GAN网络进行改进，通过相对判别器捕捉真实样本与生成样本的相对差异，缓解模式坍塌；引入R₁，R₂零中心梯度惩罚机制，保证模型局部收敛性，提升训练稳定性；在生成器中嵌入残差块，强化小样本特征提取能力，提升生成图像清晰度。实验结果表明，该方法用于MNIST及耐张线夹X光小样本数据集时，图像增强效果FID值对比未经改进的网络分别降低24.3%与29.7%。

Abstract: In response to the scarcity of deep learning image datasets in small-sample scenarios, and the issues of low image clarity and poor training stability in existing Generative Adversarial Networks (GAN), we propose a Relativistic pairing GAN (RpGAN) image data augmentation algorithm that integrates a relative discriminator and a zero-centered gradient penalty mechanism (R₁, R₂ regularization). This algorithm improves the GAN network by capturing the relative differences between real and generated samples through a relative discriminator, mitigating mode collapse. It introduces the R₁ and R₂ zero-centered gradient penalty mechanisms to ensure local convergence of the model and enhance training stability. Additionally, residual blocks are embedded in the generator to strengthen the ability to extract small-sample features and improve the clarity of generated images. Experimental results show that when applied to the MNIST and tension clamp X-ray small-sample datasets, the image enhancement effect shows that the FID values are reduced by 24.3% and 29.7% compared to the unimproved network, respectively.