SE PhD Final Defense of Yufan Luo
- Starts: 11:00 am on Friday, October 19, 2018
- Ends: 1:00 pm on Friday, October 19, 2018
ABSTRACT: Atomic force microscopy (AFM) is a powerful tool for imaging topography or other characteristics of sample surfaces at nanometer-scale spatial resolution by recording the interaction of a sharp probe with the surface. Dispite its excellent spatial resolution, one of the enduring challenges in AFM imaging is its poor temporal resolution relative to the rate of dynamics in many systems of interest. This has led to a large research effort on the development of high-speed AFM (HS-AFM). Most of these efforts focus on mechanical improvement and control algorithm design. This dissertation investigates a complementary HS-AFM approach based on the idea of undersampling which aims at increasing the imaging rate of the instrument by reducing the number of pixels in the sample surface that need to be acquired to create a high-quality image.
The first part of this work focuses on the reconstruction of images sub-sampled according to a scheme known as µ path patterns. These patterns consist of randomly placed short and disjoint scans and are designed specically for fast, efficient, and consistent data acquisition in AFM. We compare compressive sensing (CS) reconstruction methods with inpainting methods on recovering µ path undersampled images. The results illustrate that the reconstruction quality depends on the choice of reconstruction methods and the sample under study, with CS generally producing a superior result for samples with sparse frequency content and inpainting performing better for samples with information limited to low frequencies. Motivated by the comparison, a basis pursuit total variation (BPTV) method, combing CS and inpainting, is proposed. Based on single image reconstruction results, we also extend our analysis to the problem of multiple AFM frames, in which higher overall video reconstruction quality is achieved by pixel sharing among different frames.
The second part of the thesis considers patterns for sub-sampling in AFM. The allocation of measurements plays an important role in producing accurate reconstructions of the sample surface. We analyze the expected image reconstruction error using a greedy CS algorithm of our design, termed simplied matching pursuit (SMP), and propose a Monte Carlo-based strategy to create µ-path patterns that minimize the expected error. Because these µ path patterns involve a collection of disjoint scan paths, they require the tip of the instrument to be repeatedly lifted from and re-engaged to the surface. In many cases, the re-engagements make up a significant portion of the total data acquisition time. We therefore extend our Monte Carlo design strategy to find continuous scan patterns that minimize the reconstruction error without requiring the tip to be lifted from the surface.
For the final part of the work, we provide a hardware demonstration on a commercial AFM. We describe hardware implementation details and image a calibration grating using the proposed µ-path and continuous scan patterns. The sample surface is reconstructed from acquired data using CS and inpainting methods. The recovered image quality and achievable imaging rate are compared to full raster-scans of the sample. The experimental results show that the proposed scanning combining with reconstruction methods can produce higher image quality with less imaging time.
COMMITTEE: Advisor: Sean Andersson, SE/ME; Hua Wang, SE/ME; Roberto Tron, SE/ME; Prakash Iswhar, SE/ECE; Chair: Rebecca Khurshid, SE/ME
- 15 Saint Mary's Street, Rm 105