Edges Detection of Nanowires and Adaptively Denoising with Deep Convolutional Neural Network from SEM Images

The denoising of SEM images and extracting of image information have attracted attention in the micro-nano research fields. In this paper, we analyze the noise types and characteristics of SEM images and propose the method of deep convolutional neural network to denoise SEM images. On the basis of denoising, suitable methods are used to identify the position of single, multiple nanowires and positional relationship of multiple nanowires in the depth direction. Because of special imaging principles, there are many differences between SEM images and ordinary gray images. To the SEM image, noise has more significant influences on extracting information from SEM image. Samples which have various types and morphologies were observed under different types of electron microscopy and observation conditions to produce dataset. Then the supervised deep convolutional neural network was trained by this dataset, which has a good adaptive on denoising of SEM images. After denoising, the identification method of nanowires was further proposed which could accurately recognize single, multiple nanowires by canny edge detection and Hough transform, and quickly identify their positional relationship.