RESEARCH PROJECTS
Parallelization of Markov random field techniques
Markov Random Field (MRF) algorithms are powerful tools in image analysis to explore contextual information of data. However, the application of these methods to large data means that alternative approaches must be found to circumvent the NP-hard complexity of the MRF optimization. As part of the project "Towards Exascale: High Performance Visualization and Analytics" led by W. Bethel, we introduced a MRF-based framework that overcomes this issue by using graph partitioning (PMRF). The computational complexity is decreased as the optimization/parameter estimation are executed on small subgraphs. We have applied this technique to micro-CT datasets that can reach approximately 60GB in size. These datasets are acquired at the Advanced Light Source (ALS) by users of the Beamline 8.3.2, led by D. Parkinson. The figure above shows a sample result. Following this work, which explores a threaded, shared-memory, approach, we are working on a new distributed-memory MPI version. Because of the parallel nature of the PMRF algorithm, chunks of the original large image dataset can be processed separately in dierent computing cores with minimum communication. In doing so, we can take advantage of multi-core/many-core architectures such as Xeon processors and Xeon Phi coprocessors, as available at the LBNL National Energy Research Scientic Computing facility. The segmentation of dierent phases in microCT images is the basis to obtain measurements that map to physical properties of samples. For example, one potential solution to reduce the concentration of carbon dioxide in the atmosphere is the geologic storage of captured CO2 in underground rock formation, also known as carbon sequestration. In order to guarantee that this process is both ecient and safe, tools that provide measurements of media porosity, and permeability estimates, including visualization of pore structures are essential.
Fast 3D non-linear filtering
F3D is a package that provides accelerated 3D nonlinear lters, with OpenCL kernels that explores graphics cards technology. The Image Analysis team inside DAV, led by D. Ushizima, started this project with the development of QuantCT, a tool for classication and feature description of microCT data. F3D came to overcome issues related to QuantCT, specially in terms of scalability and exibility in applying 3D lters to huge amounts of data. My contribution, among others, falls on the communication of the Java code with the OpenCL kernels, the ability to provide in-memory streaming (apply sequence of lters), and the scripting feature. F3D can be easily applied to dierent science problems. One example is the work we have been conducting in collaboration with the ALS, General Electric, Air Force, and University of Utah for the analysis of ceramic composites, which are being used for the construction of next generation jet engines. The initial point of this project was the work developed by H. A. Bale and R. O. Ritchie who created a new acquisition equipment that provides the capability of obtaining microCT images of samples under dierent environment conditions such as change of temperature and application of loads. Given the images obtained from these experiments, the challenge is how to extract relevant information and measurements from the data in order to assess the material. For example, in our recent works, we combine F3D with machine learning, specically template matching, to detect structures in 3D images, including bers and cracks. Taking advantage of a model that minimizes data movement and uses intra-socket parallelism, F3D runs tens of times faster than traditional techniques. Results show the viability and accuracy of the methods for the assessment of microCT images of material samples. The figure above shows the user interface of the plugin and a sample result.
Please access the following links for more details:
Structure recognition from high resolution images of ceramic composites
Fast detection of material deformation through structural dissimilarity
​
Streaming framework for real-time ptychographic reconstruction
Ptychography is an imaging modality that enables one to build up very large images at wavelength resolution (i.e. potentially atomic) by combining the large eld of view of a high precision scanning microscope system with the resolution enabled by diraction measurements. The team involving D. Shapiro, leader of the ALS Beamline 5.3.2.1, and S. Marchesini is making important advances developing diractive imaging methods achieving spatial resolutions never seen before. However, there are several software/algorithmic related challenges involved with the image acquisition system and the ptychographic image reconstruction. When I started working with this team, including H. Krishnam (also member of the Image Analysis team of the DAV group), there was an important challenge to tackle: the development of a real-time streaming pipeline for ptychographic image acquisition and reconstruction. In collaboration with the Uppsala University - Sweden, we are developing a software infrastructure for realtime streaming and ptychographic reconstruction. There are two main parts involved in this framework: the graphical user interface (GUI) and the processing modules (backend). My contributions fall in both parts: (1) The development of the GUI is essential for the streaming pipeline, targeting mainly the control of the experiment by the user, real-time feedback and visualization. (2) In the backend side, I am responsible for the development of the dierent processing modules and the network communication between them. I am also involved in the rewriting process of the reconstruction algorithm, required to adapt it for the streaming architecture, and the parallel processing scheme. Besides overcoming the series of I/O operations and the complexity of the common pipeline used at the experimental facilities, the new pipeline provides real-time feedback while the experimental acquisition is in process. Figure 3 shows the user interface provided with the real-time streaming infrastructure. After setting the parameters needed for the acquisition and reconstruction process, the user triggers the streaming pipeline. As data come from the acquisition equipment through the pipeline, the user interface is updated with the result of the ptychographic reconstruction, along with the STXM image (Scanning Transmission X-ray Microscopy), illumination and diraction patterns. The infrastructure works so that the process is recorded and the user can stop and restart it when necessary, and also save the nal reconstruction result for future analysis.
Please access the following link for more details:
​
Analysis of films for microelectronics
Controlling a material's porosity is useful in revealing and netuning its properties as a dielectric, sorbent, or active layer for applications in catalysis, health, and energy. Pores with mesoscale dimensions are of particular interest because they can be introduced by embedding molecular or polymeric porogens within the host material and then processing the composite to create mesopores. The analysis of ordered architectures for a wide range of porosity is essential to understand how mesopore dimensions, shape, and spatial arrangement dictate the properties of porogen packing. To advance the science of mesoscale assembly, a careful revaluation of the factors governing porogen packing and shape persistence before and after processing is needed. The analysis of scanning transmission electron microscopy (STEM) tomography images plays an important role in understanding porogen packing. Scientists at the National Center for Electron Microscopy (NCEM) needed tools to tackle this problem. In collaboration with P. Ercius, we started to develop image processing techniques to help on the analysis and validation of lms for microelectronics. In this paper, we describe a framework to understand the fundamental packing limits for spherical block copolymer (BCP) micellar porogens during the assembly and thermal processing of periodic mesoporous organosilicas (PMOs). Part of this framework consists of analyzing STEM tomography images, which can provide information about both ordered and disordered domains in 3D space. We were able to assess pore architecture through image texture analysis as shown in the figure above. This project was conducted in collaboration with Intel for the development of more ecient microchips.
Please access the following link for more details:
​
Doctorate's Thesis
Markov random field framework for the segmentation of thin and ramified structures
My thesis work (at University of Sao Paulo with Professors R. M. Cesar Jr., R. Hirata Jr.) focused on exploring and developing techniques for the segmentation of images containing thin, elongated and ramied structures. This type of shape is present in several dierent scientic elds. In medical imaging for example, the detection of arteries, capillaries, veins and neurons in images is essential for the diagnosis and treatment of life threatening diseases. In the geoscience eld, the accurate recognition of river and road networks is of major signicance for applications such as urban planning, maps, trac management, and cartography. Several diculties are involved in the process of detecting this kind of structure. Their spectral and spatial characteristics are usually very complex and variable. Additionally, the thinner ones are very "fragile" to any kind of image processing techniques applied to the image, making the loss of structure easy. Another very common challenge is occlusion of part of the structures, either because of low image resolution or due to image acquisition problems. Consequently, the process of extracting these structures from images is a constant challenge when developing any image analysis pipeline. In this sense, I have developed a framework that combines a Markovian model and high level perceptual concepts of computer vision (Gestalt perceptual laws), along with multi-modality image fusion, achieving better topological and perceptual representation of the structures leading to better extraction results. Part of this work was developed at Telecom ParisTech, Paris - France, under the supervision of F. Tupin, an expert on the eld of Markov Random Fields applied to satellite images. Specically, the core development of the Markovian framework and the application of the technique to the detection of roads and rivers were executed during my work in France, with the collaboration of the Image Processing and understanding Group. I also went to work at the Indian Statistical Institute in Kolkata and the Indian Institute of Technology in Kharagpur under another international collaboration, where the developed approach was also applied to the detection of blood vessels in retinal images, aiming the diagnosis of Diabetic Retinopathy.
Please access the following links for more details:
How to combine TerraSAR-X and COSMO-Skymed high-resolution images for a better scene understanding?
Parameter Estimation for Ridge Detection in Images with Thin Structures
Ridge Linking Using an Adaptive Oriented Mask Applied to Plant Root Images with Thin Structures
Master's Thesis
Deconvolution of vibroacoustic images
This project aimed the quality improvement of the images obtained by an acquisition system called vibro-acoustography. I started working with this technique during my Master's thesis project conducted with Prof. N. D. A. Mascarenhas at Federal University of Sao Carlos (Brazil), along with Professors A. C. Frery and G. T. Silva, both from the Federal University of Alagoas (Brazil), and collaborators from the Mayo Clinic (Rochester, United States) Professors M. Fatemi and M. Urban. Vibro-acoustography, an imaging modality based on ultrasound-stimulated acoustic emission, has been successfully used in several medical applications. The main objective of my Master's project was to explore the image formation process of this acquisition modality and develop image restoration methods to improve the quality of the images. The importance of this research is that in almost all cases this technique can only obtain images in a unique plan (C-scan), because the degradation is very high in the axial direction. We tackle this problem using image restoration techniques, enabling the capability of acquisition in all plans, i.e., generating 3D vibro-acoustic images. This new feature along with the high resolution, contrast and signal-to-noise ratio, has direct impact on detecting early stages of pregnancy related diseases, cancers and injuries. Furthermore, the use of this same imaging technique can take the inspection of defects and flaws in materials such as carbon-fiber reinforced plates and steel-concrete composites to a much higher precision level.
Please access the following link for more details:
​