Abstract: To solve the problems of mode collapse,image blur and misjudgment in the deep learning of small sample data set of strip steel,an improved Auxiliary Classifier Generative Adversarial Networks (ACGAN) data enhancement method was proposed.The Wasserstein distance with gradient penalty term was used as the loss function to solve the problem of pattern collapse and training instability.The label deconvolution network was improved in the generator network to make the label information run through the whole generator network better.A denoising structure was designed at the end of the generator network to improve the quality of the generated image.Then,the idea of cascade fusion was introduced into the discriminator network to enhance the ability of network discrimination.Finally,the improved model was compared with the NEU steel strip surface defect data set and MNIST data set.The results showed that the definition and accuracy of images generated by the proposed model were improved obviously,and the mean value of objective index FID was decreased by 15.8% in the NEU steel strip surface defect data set and 73% in the MNIST data set.It provided a new method for the expansion of steel strip small sample data set.

Key words: image production, generative adversarial networks, data enhancement, small sample

摘要： 为了解决带钢小样本数据集在深度学习中出现的模式崩溃、图像模糊和错判等问题,提出一种改进的ACGAN数据增强方法。首先,模型中引入带梯度惩罚项的Wasserstein距离作为损失函数,解决了模式崩溃和训练不稳定问题;其次,生成器网络中改进标签反卷积网络,使标签信息更好地贯穿整个生成网络,并在其末端设计了去噪结构,提高了生成图像质量;接着,判别器网络中引入级联融合思想,增强了网络判别能力;最后,将改进前后的模型在NEU带钢表面缺陷数据集和MNIST数据集上进行对比实验,结果表明:所提模型生成各类样本图像的清晰度、准确性明显提高,并且客观指标FID的平均值在NEU带钢表面缺陷数据集上下降了15.8%,在MNIST数据集下降了73%,为带钢小样本数据集的扩充提供了一种新方法。

关键词: 图像生成, 生成对抗网络, 数据增强, 小样本