A Comparison of Image Quality Evaluation Techniques for Transmission X-Ray Microscopy

Bolgert, Peter J ; /Marquette U. /SLAC

关键词: ALGORITHMS;    COMPUTERS;    DESIGN;    EVALUATION;    IMAGE PROCESSING;    METRICS;    MICROSCOPES;    MICROSCOPY;    RESOLUTION;    SHAPE;    SYNCHROTRON RADIATION OTHER;   

DOI  :  10.2172/1049731
RP-ID  :  SLAC-TN-12-018
PID  :  OSTI ID: 1049731
Others  :  TRN: US201218%%596

学科分类：社会科学、人文和艺术（综合）

美国|英语

来源: SciTech Connect

【 摘 要 】

Beamline 6-2c at Stanford Synchrotron Radiation Lightsource (SSRL) is capable of Transmission X-ray Microscopy (TXM) at 30 nm resolution. Raw images from the microscope must undergo extensive image processing before publication. Since typical data sets normally contain thousands of images, it is necessary to automate the image processing workflow as much as possible, particularly for the aligning and averaging of similar images. Currently we align images using the 'phase correlation' algorithm, which calculates the relative offset of two images by multiplying them in the frequency domain. For images containing high frequency noise, this algorithm will align noise with noise, resulting in a blurry average. To remedy this we multiply the images by a Gaussian function in the frequency domain, so that the algorithm ignores the high frequency noise while properly aligning the features of interest (FOI). The shape of the Gaussian is manually tuned by the user until the resulting average image is sharpest. To automatically optimize this process, it is necessary for the computer to evaluate the quality of the average image by quantifying its sharpness. In our research we explored two image sharpness metrics, the variance method and the frequency threshold method. The variance method uses the variance of the image as an indicator of sharpness while the frequency threshold method sums up the power in a specific frequency band. These metrics were tested on a variety of test images, containing both real and artificial noise. To apply these sharpness metrics, we designed and built a MATLAB graphical user interface (GUI) called 'Blur Master.' We found that it is possible for blurry images to have a large variance if they contain high amounts of noise. On the other hand, we found the frequency method to be quite reliable, although it is necessary to manually choose suitable limits for the frequency band. Further research must be performed to design an algorithm which automatically selects these parameters.