基于改进DCGAN的铸件表面缺陷生成方法

Casting surface defect generation method based on improved DCGAN

为解决传统数据增强仅通过改变和翻转图像无法生成新的铸件表面缺陷图像且泛化能力较差等问题,通过修改网络结构并将注意力机制嵌入深度卷积生成对抗网络DCGAN模型中,实现更丰富的特征聚合,加强关键信息的交互,同时用MetaAconc激活函数替代ReLU激活函数,提升网络的自适应激活能力,增强模型训练的稳定性,并提出AMDCGAN模型。实验结果表明,相比DCGAN模型,AMDCGAN模型生成图像的FID值从19.07降至8.06,能更有效地完成铸件表面缺陷图像的生成任务。

In order to solve the problem that traditional data enhancement cannot generate new casting surface defect images only by changing and flipping images and lack of good generalization ability,richer feature aggregation is achieved and the interaction of key information is strengthened by modifying the network structure and embedding the attention mechanism into the deep convolution generative adversarial network DCGAN model.At the same time,the MetaAconc activation function is used to replace the ReLU activation function to improve the adaptive activation ability of the network,and the stability of model training,and the AMDCGAN model is proposed.The experimental results show that compared with DCGAN,the AMDCGAN generation image FID is reduced from 19.07 to 8.06,which can more effectively perform the casting surface defect generation task.