Summer 2009 REU Image Processing Projects

Problem: What is the best dimension reduction method for hyperspectral imagery?

While a standard color image contains 3 bands for Red Green Blue (RGB) light, a hyperspectral image typically contains about 200 bands. Each band is a grayscale image representing the sensor response to a particular wavelength of light. Often the relevant information in the image can be described in a much smaller number of image bands. That is, a 200 band image can often be compressed to a 4 band image that still describes the important image classes (tree, grass, bulding, etc.). Many different dimension reduction (DR) methods have been developed. These methods include classical linear methods such as PCA, ICA, and MDA. Nonlinear methods such as ISOMAP, Diffusion Maps, and LLE have proven to be more effective in reducing the dimension while still preserving the image information, but these methods are much slower and generally have large memory requirements. In 2008, researchers developed hybrid linear methods such as GPCA that represent a compromise between the speed of the linear methods and the effectiveness of the nonlinear methods. To our knowledge, hybrid linear methods have not been tested on hyperspectral data.

Despite the wide variety of DR methods and the availability of code, there has been little work done in comparing these methods and deciding which DR method is best in a given situation. A student team will test the effect of DR methods on the three most important tasks in hyperspectral image processing: classification, anomaly/target detection, and linear demixing. Students will be expected to gather data sets, in some cases preparing synthetic images for testing purposes. A major part of this project will involve developing metrics and numerical experiments for comparing the performance of the DR methods. In other words, the students will be asked to determine the best DR method for hyperspectral images, but the hard part is defining what is meant by the word "best".

Prerequisites: Students should have a background in linear algebra. Some background in statistics and numerical analysis would be helpful, but is not required.

Problem: How do we register two images from different modalities?

There are many applications when two images need to be lined up. For example, suppose we want to evaluate the effect of global warming by comparing the ice shelf over the Arctic today against a picture of the Arctic ten years ago. Before we can compute the difference, we first need to accurately align the images, a process called "image registration". Image registration is a necessary first step for many tasks such as change detection, image fusion, and pan-sharpening. In general, image registration is a very difficult problem. In our example of Arctic images, we want to register the images so we can detect the differences, but it is difficult to align the images because they are different!

Many image registration methods have been developed, but the choice of registration method appears to be application dependent. A method developed for registering MRI brain images may not be appropriate for registering satellite images of the earth. Focusing on data gathered from earth-observing satellites (like pictures on Google Maps), students will test a variety of image registration methods. The students will investigate both rigid and non-rigid registration techniques as well as both area-based and feature-based approaches. Possible methods include normalized cross correlation, variational methods, wavelet-based methods, and feature matching. Students will examine existing software for image registration such as the Matlab central library, ITK, and AIR. A particular area of interest is to register images from different modalities, such as a color image and a hyperspectral image. This topic is of particular interest to faculty at UCLA working on image fusion.

Prerequisites: Students should have some experience with computer programming. Some background in image processing would be helpful, but is not required.

Problem: Can we guide the path of a needle on an atomic force microscope to reduce the imaging time?

Atomic Force Microscopy (AFM) is a revolutionary technology which allows researchers to view materials at the molecular level. Scientists use AFM to study the structure of materials and to observe processes such as the binding of proteins. But AFM doesn't work like a camera that takes snapshot of a scene in one flash. AFM more closely resembles a record player, where a tiny needle called the cantilever gently rides the material surface, recording the precise height as it moves. Typically, the cantilever moves back and forth across the material row by row until an entire image is scratched out. Gathering one image typically takes 10 seconds. This is acceptable when the material being studied is static, but in some applications the material is in motion. For example, biologists would like to observe the microscale binding of proteins, a process which lasts less than 1 second. So by the time the cantilever has arrived at the protein location, the process we were trying to observe has already finished! Rather than gathering an image of the entire material surface, an alternative would be to localize the needle over the protein, in effect "cropping" the image to a specific area.

A student team will investigate techniques for guiding the cantilever to a region of interest, with the ultimate goal of reducing the time taken to gather the image. Using ideas from boundary tracking and pattern recognition, students will develop a simulator in Matlab that shows how the cantilever path can be restricted to a region. Note that the region of interest depends on the application, so the students will examine data from different sources such as the crystalline structure of semiconductors and the intertaction of proteins. Although we do not have access to an AFM for the students to directly program, the students will demonstrate their "proof of concept" in two ways: software with a graphical user interface and a robotic simulation.

Prerequisites: Students should have some experience with computer programming. Some background in image processing and robotics would be helpful, but is not required.

Problem: Can we automatically extract the blood vessel network from the image of a human placenta?

Recent medical research indicates that the blood vessel network in a human placenta can give information about the fetal origins of adult health risk. Extracting the network from a 2D image can be difficult due to the lack of contrast and other visual cues as well as the three-dimensional nature of the circulatory network. Using hand-labeled placenta images provided by Placental Analytics LLC as a training dataset, a student team will develop an algorithm to mimic the human operator and automatically extract the blood vessel network from an image. The students will explore different machine learning algorithms, including neural networks and multivariate regressions. The goal will be to generate a list of coordinates (x,y,w) indicating the position and width of the blood vessels in a given image. Time permitting, the students will investigate performance metrics to benchmark their algorithms and useful statistics for quantifying the characteristics of the extracted network. To demonstrate their algorithms, the students will develop interactive software with a graphical user interface (GUI).

Prerequisites: Students should have some experience with computer programming. Some background in image processing and machine learning would be helpful, but is not required.