Grain Boundary Detection Of Perovskite Material Based On Machine Learning

In recent years,perovskite based solar cells,light emitting diode and detectors have been rapidly developed,and the performance of perovskite based devices could be improved by crystalline quality improvement and morphology optimization.However,the presence of grain boundaries has been proved as an important factor in disrupting the device performance.Hence,the efficient detection of those surface defects is of great importance in the performance optimization of perovskite based devices.In the traditional processing of materials,defects are mainly detected via direct observation by human being.There are many shortcomings of this traditional method including high salary of staff,slow detection efficiency,low accuracy and the deviations from individuals.In this thesis,a machine learning algorithm was utilized to study how to improve the grain boundary recognition rate of as-synthesized perovskite materials.Three kinds of grain boundary detection algorithms are proposed: the combination with traditional image feature extraction and SVM method,the VGG₁6 model method and the improved convolutional neural network model.The results are as follows:?1?Traditional image feature extraction mainly includes HSV color model,LBP local texture feature extraction and Texton texture feature extraction.After the original image is preprocessed,the SVM algorithm is used to classify the image features.Because of the need for manual feature extraction,the material characteristics are more demanding.The experimental results show that the recognition rate reached 74.86%.?2?In order to further improve the accuracy of recognition and lessen preprocessing work,this thesis designs a VGG₁6 model based on convolutional neural network to train image features.With a large number of datasets and continuous optimization of core parameters,the final recognition rate reached 89.77%.?3?The essence of the grain boundary detection work is the image classification.By improving the existing VGG₁6 model,a lower level of neural network structure is proposed to replace the original model.Under the condition of reducing the sample data set and reducing the training time of the model,the final experimental results show the improved model recognition rate can reach 94.14%.