| Keywords [eng] |
convolutional neural network, artificial neural network, YOLO, CNN, AI, YOLOv5, atomic force microscopy, object detection, voronoi entropy, objects classification, image processing, clusterization, activation function, NanoScope Analysis, voronoi diagram |
| Abstract [eng] |
The goal of this thesis is to use an artificial intelligence approach for defect detection in AFM images in order to preserve the clusterization parameters such as Voronoi entropy and Voronoi σ. The first chapters will cover the theoretical part by explaining scientific terms, describing the biological nature of tethered bilayer lipid membranes (tBLMs), defects and what's their danger to living organisms. In chapters 4 and 6, the reader will be introduced to Voronoi diagramming and the basics of digital image processing as well as methods of labeling AFM pictures. Besides that, there will be given a brief overview of feature extraction techniques such as the Hough circle transform and its performance on AFM images. But the main emphasis of this thesis is on the usage of a convolutional neural network for the detection of defects inside clusters because it may improve the prediction of electrochemical impedance spectroscopy response. A CNN model named YOLOv5 was selected as a basis for future modifications of the activation function. For the training process were prepared datasets with different combinations of channels and defect types. In the end, the model demonstrated a significant improvement over the author's research work by maintaining clusterization metrics relatively close to the original ones and improving the accuracy of defect detection within clusters. In addition, the model is able to detect both individual defects and defects within clusters, which can be considered a success. |
