基于卷积去噪自编码器的芯片表面弱缺陷检测方法

芯片表面缺陷会影响芯片的外观和性能,因此表面缺陷检测是芯片生产过程中的重要环节。具有缺陷与背景对比度低、缺陷较小等特点的弱缺陷给传统检测方法带来了挑战。因为近年来深度学习在机器视觉领域展现出了强大的能力,所以文中采用基于深度学习的方法来研究芯片表面弱缺陷的检测问题。该方法将芯片表面缺陷看作噪音,首先应用卷积去噪自编码器(Convolutional Denoising Auto-encoders,CDAE)重构无缺陷图像,然后用重构的无缺陷图像减去输入图像,获得包含缺陷信息的残差图。因为残差图中已经消除了背景的影响,所以最后可以基于残差图较容易地进行缺陷检测。由于基于CDAE重构芯片背景的无缺陷图像时存在随机噪音,导致弱缺陷可能会湮没在重构噪音中,为此,文中提出了重叠分块策略抑制重构噪音,以便更好地检测弱缺陷。因为CDAE是无监督学习网络,所以训练时无需进行大量的人工数据标注,这进一步增强了该方法的可应用性。通过对真实芯片表面数据进行测试,验证了所提方法在芯片表面检测上的有效性。

Detection Method of Chip Surface Weak Defect Based on Convolution Denoising Auto-encoders

Chip surface defects can affect the appearance and performance of the chip.Therefore,surface defect detection is an important part of the chip production process.The automatic detection method based on machine vision attracts much attention because of its advantages of low cost and high efficiency.Weak defects such as low contrast between defects and background and small defects,bring challenges to traditional detection methods.Because deep learning has shown strong capabilities in the fields of machine vision in recent years,this paper studied the detection of weak defects on the chip surface by using the method based on deep learning.Chip surface defects were regarded as noise in this menthod.Firstly,convolutional denoising auto-encoders(CDAE)is applied to reconstruct the image without defect.Then,the reconstructed image without defect is used to subtract the input image,thus obtaining the residual image with defect information.Because the influence of background has been eliminated from the residual diagram,it is easier to detect defects based on the residual diagram.Because of the random noise in the process of reconstructing defect-free image from chip background image based on CDAE,the weak defect may be lost in the reconstructed noise.Therefore,this paper proposed an overlapping block strategy to suppress the reconstructed noise,so as to better detect the weak defect.Because CDAE is an unsupervised learning network,there is no need to perform a large amount of manual data annotation during training,which further enhances the applicability of the method.By using the real chip surface data provided by the paper partner,the effectiveness of the proposed method in chip surface detection is verified.