作者:王海涛,高玉栋,侯建新,何勇军,陈德运
Authors:WANG Hai-tao,GAO Yu-dong,HOU Jian-xin,HE Yong-jun,CHEN De-yun摘要:摘要:近年来，深度学习被广泛应用于缺陷检测。目前方法可以检测较大的缺陷，但对于细微缺陷还是无法准确检测。针对这一问题，提出了一种基于深度卷积生成对抗网络(Deep convolutional generative adversarial networks，DCGAN)网络的印刷缺陷检测方法。该方法通过以下几点来提高检测精度：①在原有网络的基础上增加上采样模块，减少上采样中的损失；②提出一种自注意力机制，生成结构性更复杂和细节更准确的图像；③统计分析生成图像的噪声分布，确定最佳阈值，去除噪声，获得准确的缺陷图像。该方法加入了去噪处理，优化了网络结构，提高了DCGAN生成图像的精度。实验表明，与现有方法相比，在小于5像素的缺陷检测实验中，本方法可以使检测精度提高10%。
Abstract:Abstract:In recent years, deep learning has been widely used in defect detection. At present, the method can detect large defects, but it is still unable to detect the fine defects accurately. In order to solve this problem, this paper proposes a new method of printing defect detection based on deep convolutional general advanced networks (DCGAN) networks. The method improves the detection accuracy by the following points: ①The upper sampling module is added on the basis of the original network, and the loss in the upper sampling is reduced; ②A self-attention mechanism is proposed to generate more complex and accurate images; ③The noise distribution of the image is analyzed statistically, the optimal threshold is determined, the noise is removed and the accurate defect image is obtained. This method adds denoising processing, optimizes the network structure, and improves the accuracy of DCGAN image generation.The experiment shows that the accuracy of the method can be improved by 10% in the defect detection experiment to less than 5 pixels compared with the existing method

PDF全文下载地址:

可免费Download/下载PDF全文