针对小样本条件下深度学习缺陷检测算法识别率较低的问题,提出一种基于双通道生成对抗网络的数据增强方法。由全局鉴别层和局部鉴别层两通道组成生成对抗网络,其中局部鉴别器可以增加缺陷类型的置信度损失,实现对局部信息的增强。采用所提方法在镜片缺陷图像数据集上进行实验。实验结果表明,所提方法的最近邻留一指标、最大均值差异和Wasserstein距离分别达到0.52、0.15和2.81;对于麻点、划痕、气泡和异物的缺陷类型图像,生成的图像质量优于条件生成对抗网络、Wasserstein距离生成对抗网络和马尔科夫判别器。双通道生成对抗网络生成的镜片图像有着多样性的全局信息和高质量的细节特征,可以有效增强镜片缺陷数据集。

Aiming at the problem of the low recognition rate of the deep learning defect detection algorithm under the condition of small samples, a data enhancement method based on two-channel generative adversarial network is proposed. The generative adversance network is composed of two channels, such as global discriminator and local discriminator. The local discriminator can increase the confidence loss of the defect type and realize the enhancement of local information. The proposed method is used to conduct experiments on the lens defect image dataset. Experimental results show that the nearest neighbor index, maximum mean difference, and Wasserstein distance of the proposed method are 0.52, 0.15 and 2.81, respectively. For the defect type images of pitting, scratches, bubbles and foreign bodies, the generated image quality is better than that of conditional generated adversarial network, Wasserstein distance generated adversarial network and Markov discriminator. The lens image generated by the dual-channel generation confrontation network has diverse global information and high-quality detailed features, which can effectively enhance the lens defect data set.