Dalila B. Megherbi

Biography:

Dr. Dalila B. Megherbi received the Sc.M in Electrical and Computer Engineering, the Sc.M in Applied Mathematics, and the Ph.D in Electrical and Computer Engineering from Brown University, Providence, RI, USA. Dr. Megherbi is a faculty of Electrical and Computer Engineering at the University of Massachusetts, Lowell. She is the founder and Director of the research Center for Computer Machine/Human Intelligence Networking and Distributed Systems (CMINDS). Her research is internationally recognized. Since joining UMass Lowell, Dr. Megherbi holds more than a 120 refereed peer-reviewed publication articles, including in the IEEE and the prestigious Nature Biotechnology (impact factor 41.667). She holds US patent. At UMass Lowell, she has been the recipient of numerous research grants and contracts, as the primary lead principal investigator, from several federal agencies and the industry, including, DoD AFRL/WPAB, NSF, US FDA, NIH, Raytheon Air Missile Defense Systems, Xilinx Inc., Structural Dynamics Research Corporation, SUN Microsystems, Altera Inc., and Sky Computers Inc. She graduated more than 32 graduate students Ph.D. and MS students with a thesis option. She serves as associate editor and member of the editorial boards and reviewer for a dozen of journals, including IEEE transactions. She has been invited to organize/TPC/session chair, and to speak at several national and international conferences. She has been invited to serve on national and international peer review boards including, NSF, NIH, NASA, and National Science Foundation of Ireland. She has been the recipient of several research and teaching awards, including the recipient of the Best Paper Award of all conference tracks at the IEEE International Conference on Homeland Security, the recipient of the IEEE Control Systems Society CDC Best Paper Finalist Award, the recipient of the Best Paper Award at the IEEE international conference ROMA, the recipient of the Top Professor Award for Outstanding Academic Integrity Leadership and Service to the students, the recipient of several university of Massachusetts Lowell Outstanding Teaching Excellence Awards, the recipient, each year since this recognition award was initiated in 2010, of 6 UMass Lowell Annual Research and Scholarship Recognition Awards, in recognition of faculty with extensive scholarship during that year. She has been a member of an international project consortium led by the US FDA. She was invited, interviewed and quoted in the New York Times for her expertise in big data facial recognition for homeland security applications. She was contacted by major national and international news agencies for her work and expertise in big data and facial recognition for homeland security. She was invited, interviewed and featured by the Science News Radio Network for her expertise in Artificial Intelligence and Big Data technologies applied to Weapons of Mass Destruction. Her primary current research interests are in intelligent multi-sensor integration and characterization, in computational machine vision and intelligence, deep learning, “Big data” analytics, knowledge extraction/representation, and adaptive learning systems in distributed computing systems and networks, with applications to homeland security and the life sciences (high throughput meta-genomics). Her main research goal is the understanding of and building sensor-based machines that can be made to exhibit intelligence. The idea is to build intelligent machines/sensors and to understand certain aspects of human and animal biological intelligence.

Abstract:

Hyperspectral/multispectral sensing for situational awareness is of major importance for supporting government activities in many homeland security and defense applications, as well as in many civilian applications, such as target detection & identification, bio-imaging and medical field. This is due to its intrinsic ability to integrate spectral and spatial target information, such as buildings blocks-boundaries or roads. Many machine learning and object recognition algorithms display complex performances, yet not enough is known about these performance complexities. Deep learning techniques are becoming an ever-present and relatively popular part of recent approaches in imaging and vision. However, in many cases they are used on a virtually empirical basis without the needed understanding of their behavior. In particular, current deep learning techniques for Infrared & visible multispectral/hyperspectral machine vision are limited. Adding different layers and increasing the size of training data, (big data) will not be enough to shed light on these learning complexities. While various multi-spectral/hyperspectral detection/recognition algorithms have been proposed in the literature, unfortunately many of them remain in their infancy. This is mainly due to their lack of high detection recognition rates in the presence of time-varying image artifacts and conditions, even slight ones. In order to develop more accurate recognition systems, there is a primary need to identify and derive some of the causes that affect some hyperspectral multispectral target recognition accuracy rates. In this talk, I will focus on some of these causes and I present our latest findings on investigation and analysis of how and what factors may affect the recognition accuracy rate of some of the most popular and currently widely used deep-learning algorithm for automatic airborne multispectral building roof automatic detection and other applications. In particular, in this presentation, as an example, we propose and show how a “Fully” Convolutional Neural Network (FCNN)-based deep learning system with a novel Pixel Rearrangement technique results in significantly reduced computational complexity and improved accuracy than its state-of-the-art counterparts. With three NVIDIA GeForce GTX 1060 GPUs, our findings show an improvement in the performance by 0.3% in comparison to the top winning submission to the national building roof detection challenge competition that took place in year 2017, but with an additional 43% reduction in the number of deep learning architecture layers. We also present a benchmark comparison result with various existing approaches to highlight our reduced computational complexity but improved accuracy.