Recognition Algorithm of Carbon Chains in Nanoparticle Assemblies from Electron Microscopy Images

This paper presents the results of the development and application of neural networks for the automated analysis of images obtained from an electron microscope to recognize gold nanoparticles and carbyne chains. The YOLOv4 architecture, trained on 5 annotated images, was used to solve the nanoparticle detection task. The model demonstrated stable results with a loss function of 1.6, indicating its high accuracy in object localization. To detect connections between particles, the YOLOv8 architecture was applied, trained on 13 images with polygonal masks. The model exhibited high efficiency in segmenting connections, as confirmed by the Precision, Recall, and F1-score metrics. However, difficulties were encountered in detecting small particles and connections under low contrast conditions, requiring further optimization. Testing on test data confirmed the models' functionality, although cases of false positives and missed detections were noted. Future work suggests employing more advanced architectures, such as YOLO11, as well as increasing the volume and diversity of training data to enhance model accuracy and robustness. The developed solutions can be used to accelerate the analysis of nanostructures in materials science, nanoengineering, and other scientific fields.