A two-dimensional material recognition image algorithm based on deep learning

The Van der Waals heterojunction is composed of stacked atomically thin two-dimensional materials. As a new type material, it has many unprecedented characteristics. However, owing to the position and shape of the two-dimensional material peeling are random during the preparation process, the preparation efficiency of Van der Waals heterojunction is always low. Finding the peeled two-dimensional material quickly has become the key to improve the preparation speed of Van der Waals heterojunctions. In recent years, deep learning has become more and more excellent in the field of image recognition, and the application of deep learning on two-dimensional material recognition has achieved a very good effect. In this paper, we use deep learning-based semantic segmentation network to identify and segment two-dimensional materials in images collected under optical microscope. Moreover, the network architecture was improved to achieve an average accuracy of 97.17%. Since deep learning algorithms can extract high-level features of images through training with a large amount of data, compared with traditional graphics algorithms, the deep learning algorithm have good generalization capabilities and are robust to changes in conditions during detection, which can avoid errors caused by traditional algorithms that require manual adjustment of parameters.