基于生成对抗网络的色盲测试图生成

Color Blindness Test Image Generation Based on Generative Adversarial Networks

为了解决传统色盲测试图数量不足,样式单调,复用性不高等问题,提出了改进的带梯度惩罚的Wasserstein生成对抗网络(WGAN)色盲测试图生成方法。将梯度惩罚函数引入生成对抗网络(GAN),避免训练过程中的梯度消失,将重构损失、L1损失、结构相似性损失与梯度惩罚函数相结合,有效提高了色盲测试图的质量。并且加入了自注意力机制,可以更好地学习全局特征,保证了色盲测试图图案内容的多样性。生成样本的质量根据峰值信噪比(PSNR),FID(Frechet Inception Distance)及结构相似性(SSIM)进行评估,结果表明,上述模型生成的色盲测试图细节更为清晰,质量更高,优于CGAN,DCGAN,WGAN,WGAN-GP模型,具有实用意义。

In order to solve the problems of difficult produce,monotonous style and low reusability of traditional color blindness test images,an improved Wasserstein generation adversarial network(WGAN)color blindness test image generation method with gradient penalty is proposed.The gradient penalty function is used the generation adversarial network(GAN)to avoid the gradient disappearing during the training process.The reconstruction loss function,L,loss function,structural similarity loss function and gradient penalty function are combined to effectively improve the quality of color blindness test image.In addition,the use of a self-attention mechanism can provide better learning of the global features and ensure the diversity of the pattern content of the color blindness test image.The indexes of PSNR,FID,and SSIM are used to evaluate the quality of the generated samples.The results showed that the color blindness test image generated by the model was clearer in detail and higher in quality,which was superior to models of CGAN,DCGAN,WGAN,and WGAN-GP,and can be used in practice.