Abstract
The agglomeration causes a significant challenge to nanodielectrics. Identification of agglomerates in scanning electron microscopy (SEM) images is an important step of solving this issue. Motivated by the fast development of image recognition in computer vision, we propose a new approach for agglomerates identification in SEM images of nanodielectrics by semantic segmentation, which is more efficient and accurate than traditional methods. Three models based on convolutional neural networks are investigated in this work, namely pixel blocks classification network, full convolutional segmentation network employed with data augmentation and unsupervised self-encoding network. All three networks can preliminarily identify agglomerates of spherical silica-based blend polyethylene nanocomposites. The mean intersection over union (mIoU) of pixel blocks classification network is 0.843 and it takes 25 s to process an image. Full convolutional segmentation network only needs 0.059 s to process a sample, with a mIoU of 0.777. Unsupervised self-encoding network can reach a mIoU of 0.747 at a speed of 5.806 s. According to the amount of data sets, and requirements for different speed and accuracy, three kinds of networks can be flexibly selected.