Advanced sampling and reconstruction of images in Atomic Force Microscopy

Grevenkop-Castenskiold, Benjamin Højmose ; Nørgaard, Jacob Bøgeskov

4. term

Signal Processing and Computing, Master

2019

2019-06-06

44

Atomic Force Microscopy (AFM) kan lave detaljerede kort over en prøves overflade ved at føre en meget lille spids hen over den ved mange forstørrelser. Denne fremgangsmåde er tidskrævende, fordi spidsen skal bevæge sig punkt for punkt over hele overfladen for at indsamle målinger. Denne afhandling undersøger, hvordan scannetiden kan forkortes ved hjælp af billedrekonstruktion og et adaptivt målemønster. På rekonstruktionssiden afprøves Deep Image Prior (DIP), en metode som udnytter strukturen i et neuralt netværk til at genskabe billeder uden forudgående træning. På målesiden foreslås et to-trins adaptivt mønster: først en hurtig, grov skanning for at finde interessante områder, derefter en målrettet højopløsningsskanning kun i disse områder. I denne sammenhæng vurderes DIP som uegnet til AFM, fordi rekonstruktionskvaliteten er på niveau med simpel interpolation (at anslå manglende værdier ud fra nabopunkter), men med langt større beregningsomkostning. Det adaptive to-trins mønster giver derimod lovende resultater i de mest relevante dele af billedet. I gennemsnit opnås omkring 10x hastighedsforøgelse med cirka 44 dB PSNR i de relevante områder, hvilket indikerer god billedkvalitet. Gevinsten varierer dog fra billede til billede, men metoden sikrer, at den langsomme raster-skanning kun anvendes i de vigtigste områder.

Atomic Force Microscopy (AFM) creates detailed maps of a sample’s surface by moving a very small probe across it at many magnifications. This process is slow because the tip must scan point by point to collect measurements. This thesis explores ways to shorten scan time using image reconstruction and an adaptive sampling pattern. For reconstruction, we test Deep Image Prior (DIP), a method that uses the structure of an untrained neural network to rebuild images. For sampling, we propose a two-shot adaptive strategy: first a quick, coarse scan to locate regions of interest, then a targeted high-resolution scan only in those areas. In this setting, DIP was deemed infeasible for AFM because its reconstruction quality was comparable to simple interpolation (estimating missing values from neighbors) while being much more computationally expensive. In contrast, the adaptive two-shot pattern shows promising performance in the most relevant parts of the image. On average, it achieves about a 10x speedup while reaching roughly 44 dB PSNR in those regions, indicating good image quality. The speedup varies from image to image, but the method ensures that slow raster scanning is applied only to the most important areas.

[This abstract was generated with the help of AI]

AFM ; DIP ; Optimization ; Atomic ; Force ; Microscopy ; Adaptive ; Sample ; Pattern ; Sparse reconstruction ; Machine Learning ; Neural Networks ; Image Reconstruction ; Deep learning

Download
View record in AAU Student Projects