Abstract TRISO (tristructural-isotropic) fuel is a type of micro fuel particles used in high-temperature gas-cooled reactors (HTGRs). Among the quality evaluation methods for such particles, in-line phase contrast imaging technique (PCI) is more feasible for nondestructive measurement. Due to imaging hardware limitations, high noise level is a distinct feature of PCI images, and as a result, the dimensional measurement accuracy of TRISO-coated fuel particles decreases. Therefore, we propose an improved denoising hybrid model named as NL P–M model which introduces non-local theory and retains the merits of the Perona–Malik (P–M) model. The improved model is applied to numerical simulation and practical PCI images. Quantitative analysis proves that this new anisotropic diffusion model can preserve edge or texture information effectively, while ruling out noise and distinctly decreasing staircasing artifacts. Especially during the process of coating layer thickness measurement, the NL P–M model makes it easy to obtain continuous contours without noisy points or fake contour segments, thus enhancing the measurement accuracy. To address calculation complexity, a graphic processing unit (GPU) is adopted to realize the acceleration of the NL P–M denoising.