| Abstract [eng] |
Artificial neural networks are machine learning computational techniques that process information presented in a model when the relationships between input and output are complex and uneven. The use of artificial neural networks extracts information from the data presented, in this case images, that helps to analyze the data sets and adapt them for further use. Neural networks can be used to identify microparticles on surfaces, to detect bacteria in blood, or to break down objects in production lines and other situations. In the master 's thesis, convolutional neural networks are realized and applied in SEM experimental images. Aforementioned images contain nanoparticles that are intended to be segmented with the help of networks. These nanostructures were captured on the surfaces of annealed TiO2 thin films, which were captured under a digital electron microscope. A mask image was created for existing nanoparticles in the dataset, with each nano-particle mask marked separately. The obtained images are convenient to present in the neural network and to train it. The convolutional neural network uses the U-Net architecture, which is adapted to two different research tasks. One study segment nanoparticles, and other studies segment a couple of classes: nanoparticle, its edge, and background. This architecture helps the convolutional neural network cope with the noise in the images. Applying only a global threshold value recognizes a lot of noise in the images or loses a lot of useful information when the threshold is changed. Two methods are compared, classifying three classes, and classifying the nanoparticle, but the water-basin method is also applied. Using the latter method, the model separates the contacting nanoparticles, but sometimes does not segment the smaller particles. The segmentation of the three classes, meanwhile, yielded exceptionally satisfactory results with an accuracy of 94%. This study showed that the structure of neural networks with additional modification is easily adapted for image recognition improvement and analysis. |
