A discussion with Juan Herrera, Perceptics
VSD: Can you provide some background on your company?
Herrera: Perceptics was recently divested from Northrop Grumman and focuses on “mission-critical imaging technology.” We view the national border security segment as the primary focus of the company and have been supporting the national border security market for more than 28 years. Our License Plate Reader systems are unique in that they can identify not only the alphanumeric characters but also the state, province, or country of origin. More recently, we developed a high-resolution color imaging system for capturing and displaying images of the undercarriage of vehicles.
I’m most interested in real-time unconstrained character-recognition applications-what we like to call “extreme character recognition.” That includes reading license plates for real-time threat assessment, identifying the ISO container code that uniquely identifies every intermodal shipping container, and associating that data with other images and sensor information. In addition to being robust, fast, and accurate, the character-recognition algorithms have to be font independent, and rotation and scale tolerant.
VSD: What algorithms and specific software developments do you see emerging in the next five years?
Herrera: I think new video-compression algorithms will take full advantage of broadband network bandwidth and allow real-time transmission of live video over a network. I see the pairing of machine learning algorithms and image processing as a way of solving somewhat complex inspection tasks. Furthermore, if you monitor the results of a visual inspection system and associate them with process-control parameters, you might be able to develop a system that can automatically infer how to change those control parameters to optimize the yield.
Already, memory-based reasoning, neural networks, and fuzzy neural networks, among other algorithms, have been used to develop both supervised and unsupervised learning approaches to manage manufacturing process-control parameters and improve yield. Several authors have reported on the application of hybrid learning approaches to the semiconductor manufacturing process, where thousands of control parameters are involved in the production of each wafer and traditional statistical methods are not always effective. A typical outcome of using machine learning techniques to monitor control parameters is the gaining of additional insight into the manufacturing process.
VSD: What hardware developments do you see emerging in the next five years?
Herrera: I expect to see the development of higher-depth imagers that exhibit higher dynamic range and are less prone to blooming and streaking. Also, I expect to see LED illuminators in a wider range of light frequencies and/or with a frequency response that can be tuned over some range, and high resolution, true-color imagers where the color information for each pixel is captured in the same physical location on the imager-as opposed to being derived from three separate but closely spaced imaging elements.
Last year, Foveon (Santa Clara, CA, USA; www.foveon.com), founded by Carver Mead, released a 14.1-Mpixel image sensor that uses its proprietary X3 technology to stack red, green and blue pixels vertically on the same imager. Imagers based on the X3 technology have already been integrated into several consumer digital cameras and a couple of industrial-grade cameras for scientific, medical, security, cinema, and communications applications.
High-dynamic-range imagers (12-bit monochrome, 30-bit color), coupled with high-brightness LEDs, can help cope with the wide range of illumination conditions that occur in outdoor imaging applications. They also can help address what is probably the most challenging difficulty encountered on every machine-vision problem: image segmentation.
VSD: How do you think these developments will impact future generations of image-processing and machine-vision systems?
Herrera: Advances in illumination and image sensing will lead to systems and solutions that address real-world applications such as security, surveillance, and law enforcement that need to operate 24/7 and are exposed to the elements in a variety of ambient lighting and weather conditions. Better images, coupled with faster processors, will lead to even more complex applications. As it is, there has been an explosion of video-based “behavior sensing” software to automate the detection of “suspicious activities” for security and “event recording” applications.
One practical application of “smart imaging sensors” is the use of infrared thermography to screen airport passengers for avian flu. Using self-calibrating software and infrared cameras capable of measuring very small changes in skin surface temperature, ThermaCam from FLIR Systems (North Billerica, MA, USA; www.flirthermography.com) allows nonmedically trained personnel to quickly scan large numbers of people for symptoms of the disease. Based on the volume of work on facial recognition, it seems plausible that someone could develop a system capable of recognizing facial expressions.
VSD: When do you see biologically inspired models of the human visual system and the human cognitive process emerging, and how do you think these will be implemented?
Herrera: I don’t know that that is a necessary precondition in the progression to impart higher cognitive capabilities to machine-vision systems. After all, airplanes don’t have to flap their wings to create lift. But I do think that there are several applications in the machine-vision realm that can benefit from allowing a machine to be trained simply by “observing” how an experienced human operator performs an inspection task to automatically “learn” some basic rules.
An interesting application in this area is the work led by Michalski and colleagues at the Machine Learning and Inference Laboratory of George Mason University, where they are using inductive learning algorithms on x-ray images of luggage to detect blasting caps. At Perceptics we have developed our own set of inductive learning algorithms that we have used very successfully to generate very accurate classifiers for identifying the alphanumeric numbers and the state, province or country of origin of license plates of vehicles traveling through a checkpoint. The challenge is how to automate the learning process so that it can take place without compromising the overall stability of the decision-making system.
There is some higher-level learning process that allows humans to recognize objects that they have seen only once before. I don’t think that process can be modeled using a neural approach. Instead, the ability to build and maintain abstract models based on information extracted from image data can lead to practical solutions to relatively complex problems.
VSD: Which other wavelengths do you see becoming increasingly important in the next five years and why?
Herrera: For machine inspection, I think ultraviolet (UV) has been largely ignored. UV-sensitive imagers and lenses are relatively expensive and hard to get. However, some surface defects are easier to detect using illumination in the UV range.
For medical imaging and security applications, terahertz radiation looks the most promising due to its interesting properties and behavior with respect to living tissue and its ability to penetrate clothing, paper, and plastics. An immediate use for terahertz could be security screening at airports and sensitive facilities, where its ability to image concealed weapons through clothing and its potential for detecting plastic explosives could prove very useful.
VSD: Which areas present the most opportunities for engineers involved in machine vision or image processing?
Herrera: Ever since 9/11/2001, all applications dealing with security, such as law enforcement, protection, crime prevention, surveillance, and forensic analysis, have incorporated video images into the volume of data that is continuously captured, analyzed, and stored. I think this is a vast field with lots of potentially profitable opportunities.
Juan Herrera is technical director at Perceptics, Knoxville, TN, USA; www.perceptics.com. He has a Ph.D. in computer engineering from the University of Tennessee and more than 20 years of experience in optical character recognition, machine vision, artificial intelligence, image processing, and pattern recognition. Editor in chief Conard Holton spoke to him about trends in security and character recognition.
