一种改进DCGANs网络的磁瓦缺陷图像生成方法研究

Research on Generation Method of Magnetic Tile Defect Image Based on Improved DCGANs Network

基于机器视觉的磁瓦表面缺陷检测研究对于改进磁瓦生产工艺、提升磁瓦生产效率有着重要意义.但在研究过程中,存在磁瓦含缺陷样本收集困难、不同缺陷样本数不均匀、缺陷类型单一等问题.本文提出一种使用高斯混合模型的深度卷积生成对抗网络(Gaussian Mixture Model Deep Convolution Generative Adversarial Networks,GMM-DCGANs)生成含缺陷磁瓦图像的方法.在深度卷积生成对抗网络的基础上,将生成图像的输入噪声潜在空间复杂化为高斯混合模型,从而提高图像生成网络对有限数量且具有类间及类内多样性训练样本的学习能力.实验结果表明,GMMDCGANs网络可以生成质量更好、缺陷类型更加丰富的磁瓦缺陷图像,并且生成的图像满足缺陷检测及分类的要求.

Machine vision based surface defect detection and classificationof the magnetic tiles has been a great significance to improve the production process and the production efficiency of the magnetic tile.However,in the existing researches,there are problems such as difficulty in collecting defect samples of the magnetic tiles,uneven quantity of the samples with different defects,and single type of the defects.This paper proposes a method of Gaussian mixture model baseddeep convolution generative adversarial networks(GMMDCGANs) for generating the images of the magnetic tiles with the defects.Based on the deep convolution generative adversarial networks,the input noise potential space of the generated image is complicated into the Gaussian mixture model,thereby improving the image generation network’s learningability with a limited quantity of training samples with inter-and intra-class diversity.The experimental results show that the GMM-DCGANs network can generate the magnetic tile images with the defect by obtaining higher quality and richer modes in the existing defect types,and the generated images meet the requirements of defect detection and classification.