Ultrasmall gold nanoparticles (diameter 1 to 2 nm) can be used in medicine as drug delivery systems. They can also be functionalized to interact with proteins. Their biological properties and the degree of functionalization critically depends on their structure and size. Consequently, more information on the size of the particles and the synthetic parameters that influence their structure is essential. Different methods are available to analyze ultrasmall nanoparticles, with high-resolution transmission electron microscopy (HRTEM) being the most prominent. With HRTEM, the particle size and the internal crystal structure of the particles can be visualized. However, analyzing HRTEM images is time-consuming as it must be done manually. To effectively use HRTEM for a large-scale analysis of ultrasmall nanoparticles, an automatized image processing is desired. As a first step in image processing, an algorithm is needed to distinguish between the crystalline particles and the amorphous background. Common filters can be used to make it easier to distinguish between particles and background but are unable to reliably remove the background from the images, so that manual correction is still needed. A promising method for automatically finding the particles in a given HRTEM image is the use of semantic segmentation. A pre-existing neural network was trained on HRTEM images of ultrasmall and regular gold nanoparticles, and the network performance was analyzed.