Human beings are visually-oriented in their daily life; they use the sense of sight more than any of the other senses with which they have been endowed. The modern understanding of human vision and the underlining principles were only discovered in the past couple centuries. The nineteenth and twentieth centuries witnessed the development of photographic and digital imaging camera systems, which partially mimic human visual systems. We will open a small window on the history of human vision and camera systems, and try to compare today’s state-of-the-art cameras with the human visual system, focusing mainly on solid-state image sensors, or camera chips, and the image-sensing element of the human visual system, the eye.

History of human vision

Human vision has been the subject of conflicting interpretations since ancient times. Many ancient physicians and philosophers believed in the theory of extramission, or the active eye. According to this theory, the eye perceives objects by emanating light and seizing objects with its rays. It was in medieval Islamic culture that research on human vision and optics developed into a system similar to the modern theory of vision. Among others, Ibn Al-Haytham (Alhazen) (965-1040 A.D.), a Muslim physicist, astronomer, and mathematician in the tenth century, played a great part in this field by promoting the intromission theory which states that vision only occurs because of light rays entering the eye. Ibn Al-Haytham founded physiological optics, which distinguished the functioning of the eye from the behavior of light. On the other hand, ten centuries after Ibn Al-Haytham, Winer et al. (2002) have found recent evidence that as many as 50% of American college students believe in the extramission theory.1

Although the fundamental features, anatomy, and physiology of the eye were documented by Galen (129–200 A.D.), an ancient Greek physician, in the second century A.D., it was Kepler, a close reader of Ibn Al-Haytham, who offered the first theory of the retinal image and the correct operation of the eye in 1604. He proclaimed, “Therefore vision occurs through a picture of the visible things on the white, concave surface of the retina.” Progress came slowly after Kepler, because little was known about the nervous system until the nineteenth century, and only recently have scientists acquired a more knowledge about how the brain apprehends the retinal image. But many questions still elude us.

History of the camera

In parallel with curiosity about human vision, human beings have also tried to mimic human vision by capturing images of objects with instruments. Around 1000 A.D., Ibn Al-Haytham, also known as the father of modern optics, invented the pinhole camera,2 and explained why the image was upside down. It was Johannes Kepler who further suggested the use of a lens to improve the pinhole camera in the 1600s. Capturing an image on a photographic plate was first achieved in the early 1800s. Consequently, photographic cameras began to be mass-marketed in the twentieth century. The photographic equipment with which we are all familiar today, such as the 35mm camera, flash bulb, Polaroid camera, and the point-and-shoot auto focus camera, all were developed in the twentieth century. The invention of the camera as we know it today paved the way for other technologies, including the moving image capture, and, later, the digital camera, in which electronic image-capture devices were used. In 1972, chemically processing an image onto photographic paper no longer became the sole destination of an image, because the first filmless electronic camera was patented by Texas Instruments Corporation. Filmless electronic cameras were made possible with the invention of solid-state image-capture devices called charge coupled devices (CCD) and metal-oxide-semiconductor (MOS) image sensors in the late 1960s. Since the invention of solid-state imagers, people have become more visually stimulated and oriented than ever before in history.

A comparison of camera chips and the human eye

The technological advancements of solid-state image-capture camera chip design and manufacturing during the past twenty-five years has made digital imaging more affordable and accessible to the general public. These advancements have become more visible to consumers in mobile products, particularly in cellular phones, in which there are still and video-camera functions. Although digital cameras are easily available today, the state-of-the-art image sensor chips used in these cameras exhibit a performance gap when compared with the capabilities of the human eye. How good these image sensor chips are today when compared to our eyes is a question that will be elaborated on.

It is possible to compare the capabilities of the human eye and state-of-the-art image sensor chips used in cellular phones or in mainstream PC and digital still cameras. It is also possible to compare the capabilities of the human visual system, including the eyes, the optic nerve, the visual cortex, etc. with a digital camera system which includes optics, image-capture and signal-processing chips and other camera apparatuses. The capabilities include ability to see different colors (spectral response), photo-element (pixel) characteristics (size, density, distribution), light sensitivity, light-intensity response range, functionality and operation modes, and signal processing capabilities.

Spectral response

A single light-sensing element in a solid-state image sensor is called a pixel. In the human eye it is called the photoreceptor. Both elements convert impinging light or photons into electrical signals. The human eye sees in the so-called visible spectrum, between 380nm (blue) and 750 nm (red), and utilizes two kinds of photoreceptors on the retina; rods and cones. The cones are used for color and daylight vision. Rods are responsible for night vision. There are three types of cone photoreceptors on the retina that contain different types of photosensitive pigments. The three types of cones are L, M, and S, and they have pigments that respond best to wavelengths of light that are long or red (peak at 564 nm), medium or green (peak at 534 nm), and short or blue (peak at 420 nm), respectively. The rods (R) are most sensitive at a wavelength of approximately 498 nm (green), as seen in Figure 1.3 Image sensor pixels in digital cameras mimic the photoreceptors in the human eye for color vision. They utilize three kinds of color filters (red, green, blue) on top of each pixel to convert light rays into electrical signals in different visible spectrums. Unlike the cones in the human eye, camera pixels and color filters can be designed to cover wide spectrums that are not visible to the human eye, for instance, the x-ray, ultraviolet, and infrared spectrums. In the category of spectral response range, camera pixels exhibit greater flexibility than those of the photoreceptors of the human eye. On the other hand, interestingly enough, the eyesight that humans possess has similar spectral characteristics as the sun. The solar light emission peaks in the visible spectrum as seen in Figure 2.4

Figure 1. Spectral absorption curves of the short (S), medium (M), and long (L) wavelength pigments in human cone and rod cells.3

Figure 2. The daylight solar spectral power distribution on earth.4

Pixel and array size

The size of pixels in today’s modern digital cameras is getting closer to the size of the photoreceptors in human eye. The typical human eye contains an average of 130 million photoreceptors. The diameter of the rods and cones varies between 1.0m and 8.0m, depending on their location on the retina.5 Today’s state-of-the-art image sensor chips contain 10 to 30 million pixels. Each pixel can be as small as 1.4m in diameter. To date there has been no image sensor that is 1.4m pixel in size or more than 8 million pixels. However, the human being has been equipped with photoreceptors that are as small as 1.0m and has more than 100 million photoreceptors; and this is since the beginning of existence. It is also estimated that the resolution of the human eye is equivalent to an imager sensor chip of 576 million pixels with a 120 degree field of view.6 Thus we still have a long way to go in improving the image-sensor pixel and array sizes used in cameras if we are to match the human eye.

Pixel distribution and formation

In the human eye the photoreceptor size and densities change, depending on their location on the retina. For example, no rods exist on the focus center of the eye, which is called the fovea. Color vision photoreceptors, which total only 10% of the eye’s photoreceptors, are located mostly on the fovea. There is an irregular distribution of photoreceptors which is unique for every human being, like a fingerprint. Yet, we all see things the same, such as colors (with the exception of people who are colorblind). In camera chips, however, pixels are arrayed regularly, in two-dimensions. As the image-processing techniques and algorithms used in camera systems are linear and do not closely mimic the signal processing that exists in the human visual system, regularly arrayed pixels are required.

Light sensitivity and response range

Although the pixel sizes in image-sensor chips are approaching the size of the photoreceptors in the human eye, camera systems are not yet close to being able to match performance in terms of light sensitivity and response range. The human visual system and photoreceptors can easily adapt to very dim and bright light, with a light-intensity response range of ten billion to one (1010:1).7 This response range goes from light conditions on a bright sunny day to dim night vision. Typically, a conventional consumer camera pixel has a light intensity response range of one thousand to one (103:1).8 In a camera system, details of a captured scene are either concealed in the dark regions or washed out by the bright light, depending on the exposure settings of the system. Thus, one could say that the human visual system works ten million times (107) more efficiently than that of consumer cameras in terms of transferring scenes into images.

Operation principle

In terms of operation principles, the photoreceptors in the human eye convert light rays into electrical signals with extremely rapid electro-chemical reactions which can detect a single photon. Typically, in the image sensor pixel of a digital camera the photoelectric effect is used to convert impinging photons into electrical charges. Electrical charges are collected and stored in each pixel during the exposure period. Collected electric charges in each pixel are amplified and converted into digital ones (logic-1) and zeros (logic-0) during image readout before the image is sent to higher processing elements, such as a personal computer, digital-still or video camera. It is possible for a single photon-counting camera to be developed. However, very special and larger pixel sizes and extra apparatuses are required to build such a camera system. Thus, we could say that it is almost impossible to build imaging pixels that have the capability and dimensions of the photoreceptors of the human eye with today’s state-of-the-art technology.

Signal processing capabilities

The captured image in the human eye is preprocessed before it is sent to the visual cortex of the brain. This preprocessing consists of a data reduction operation in which nothing is lost, with a compression ratio of 130 to 1, as only 1 million optic nerves leave each eye carrying the information from 130 million photoreceptors. This compression allows the brain to process information at a rate of 25 to 150 scenes or frames per second. Typically, every pixel in an image sensor chip is first transferred to higher processing units. A data compression method is either carried out with some loss of details in the image or the compression is never used. The transfer of frames in camera chips typically takes place sequentially, reducing the speed of the image-capture operation or frame rate. Different techniques are used to maintain a capture rate of, at most, 25 frames-per-second in camera chips. With today’s technology, image sensors that have a capture rate of one million frames per second have been proposed and can be manufactured for scientific applications.

The inherent inefficiencies of image-capture in today’s image-sensor chips are hidden by employing the limitations of the human eye. For example, solid-state image sensors have always been produced with row or column-vice uncanny stripes which are easily picked up by the human eye. However, psycho-visual experiments have shown that the human eye can only detect contrasts between two adjacent gray lines when the difference is greater than 0.5%. Thus, if a camera chip is designed to have a column to column or row to row contrast of less than 0.5%, these odd stripes would not be visible.

Conclusion

Humans are visually oriented and without a doubt, our eyes are considered to be our primary source of information. It is obvious that the human visual system is extremely complex and this complexity has fascinated human beings throughout history. Yet, the underlining principles and basic functions of human vision and the eye have only been discovered during the last two centuries. These discoveries have led research in how to mimic these functions, which has resulted in moving and still-photographic and camera equipment, and the image sensors chips used in digital cameras today. Even though human beings are only taking baby steps in fully mimicking the human eye, curiosity and scientific inquiry allows us to discover functions and features of the eye and the visual pathways that will increase our knowledge and help us to build better pixels and image sensor chips.

References

1. Winer, G. A., Cottrell, J. E., Gregg, V., Fournier, J. S., & Bica, L. A., “Fundamentally misunderstanding visual perception: Adults’ beliefs in visual emissions.” American Psychologist, 57, 417-424, 2002.

2. Ertan Salik, “Pinhole Cameras, Imaging, and The Eye” The Fountain Magazine, Issue 54, pp. 30-33, April – June 2006.

3. URL: http://en.wikipedia.org/wiki/Image:Cone-response.png

4. URL: http://www.handprint.com/HP/WCL/color3.html

5. Stefan Winkler, Digital Video Quality – Vision Models and Metrics, John-Wiley & Sons, Ltd., 2005.

6. URL: http://www.clarkvision.com/imagedetail/eye-resolution.html

7. R.C. Gonzalez and R.E. Woods, Digital Image Processing, Addison-Wesley, 1993.

8. M. Schanz, et al., “A high-dynamic-range CMOS image sensor for automotive applications”, IEEE Journal of Solid-State Circuits, vol. 35, no. 7, pp.932-938, July 2000.

New Findings in Biology

Olfaction was long considered a uniquely mammalian trait. Scientists have disproven this by showing that many intertebrates can smell. For example, birds were thought to be unable to smell, although they have nostrils in their bills. John Audubon, a famous nineteenth-century bird artist, reached this mistaken conclusion by observing vultures confronted with a covered and an uncovered animal corpse, he concluded that they could not smell. The minute weight of the birds’ olfactory bulb consolidated this widespread misconception. Recent research shows that birds use smell when finding and distinguishing food, choosing proper nesting sites and mates, and following avian navigation routes. Ken Stager, an orinthogist at Los Angeles County Natural History Museum, used turkey vultures to disprove Audubon’s vulture experiment. Marine biologist Betsy Bang, who measured the olfactory bulbs and tissues in the brains of 151 bird species, calculated the olfactory bulb’s mass as being between 3% to 37% of the brain’s entire mass. This shows that the ratio, and not the weight, determines a bird’s ability to smell.

Other examples are as follows:

•Pigeons perceive small amounts of odorants. If their olfactory bulbs are blocked, they become lost.

•Certain seabirds (e.g., white chinned petrels) are sensitive to the chemical emitted by their main food (plankton), and so follow an olfactory path over the sea.

•European starlings smell the best region for their nesting site.

•Chickens detect inedible bugs (e.g., bright-colored bad-tasting caterpillars) through smell and sight.

•Salmon return to their hatching sites years later by using the unique olfactory memory of these sites left in their brains.

Smelling is far more developed in mammals, especially dogs and cats, which can sense parts per billion or trillion and can identify millions of different odorants. Science would benefit greatly if such abilities could be reproduced in sensors. But first, how does the human nose smell?

Perceiving Odors

Scientists divide human olfaction into steps. First, a potential odorant emits an odor’s basic elements: volatile organic compounds (VOCs). We perceive an odor when molecules are transformed into an odor by binding the receptor proteins.(1) After binding with certain types of VOCs, these receptor proteins cause depolarization. The electrical charges produce unique signals, which the epithelium’s sensory cells transmit to our neural network (axons).

These signals then are carried to a cluster of neural networks in the brain (glomeruli).(2) Ultimately, the impulse reaches the hypothalamus and describes the scent through a process of classification and identification.

Quantifying Scents

Human odor panels or gas chromatography and mass spectroscopy (GC/MS) are used to identify odors. Such quantification is problematic, however, because it is hard to quantify the VOC’s perception in the nose as a unit of odor. Quantifying mass, volume, temperature, light intensity, and the molecular concentration of a soluble substance in a solution are reasonably objective and can be measured as a multiple of a standard unit.

But a standard olfactory measure does not exist, for it varies according to time and environment. Odor concentration is expressed as a multiple of a threshold: 50% of human “sniffers” must detect-not necessarily identify-it. This threshold is defined by the American Society for Testing and Materials (ASTM), and is accepted as the absolute threshold of odor perception. It takes 5 or 10 odor units for the human panel to identify the odor. GC/MS also can identify the odor’s chemical composition.

Electronic Nose

After developments in electronic sight and hearing, scientists sought similar progress in odor perception. Research began at the University of Warwick (Coventry, England) in the 1980s. Its participants coined the term “electronic nose,” now commonly known as “e-nose.”(3) Their progress made it a commercial commodity with many applications.

E-noses have moved from being metal oxide devices, to conducting polymers, and now to laptop-size or pocket-size odor sensors. The Swiss Federal Institute of Technology (Zurich) has made one the size of a wrist-watch. However, current e-nose use is largely restricted to labs and military applications. Scientists are trying to match or surpass the human sense of smell’s accuracy and sensitivity, after which they will work on surpassing that of the canine species.

Uses and Advantages

E-noses have a wide application in agriculture. Since they can detect minute differences, an e-nose using polymer materials can determine whether a tomato is sun-ripened, picked green, or internally damaged, and whether apple juice comes from a concentrate or is authentic but pasteurized.

Volunteers often test such products. But who wants to determine if corn oil is rancid or canola oil is oxidized? E-noses, having no such “qualms,” detect changed odors in oil samples and provide far more accurate reports.

In animal science and poultry, e-noses provide detailed reports about spoiled food. Judy Arnold, a microbiologist in Athens, GA, researches food quality for the Agricultural Research Service (ARS).

In 1998, researchers discovered that e-noses can detect gases produced by spoiled poultry products. They claim that an e-nose can determine freshness, period of time in a refrigerator, and the amount of fat in white meat. Such an objective evaluation benefits poultry farmers and producers by eliminating returns of “funny-smelling” poultry. E-noses also can detect meat’s decay rate and bacteria, overall quality and freshness, the composition of mixed meat-part products (e.g., processed meat), and how long a ham has been dry-cured. Given this, the e-nose’s ability to examine a bundle of scents makes it very useful. It can perform hundreds of preliminary assessments that would occupy a chemist for months.

The military uses e-noses to detect land mines and traces of chemical-biological weapons. This is important, for over 100 million land mines litter 62 war-torn countries. Although dog-sniffers are useful, this practice is inhumane (dogs are often injured) and impractical (they need lots of training).

E-noses also are better than metal detectors and ground-penetrating radar and infrared imaging-the former detects even tiny pieces of metal, whereas the latter often images pebbles. As e-noses can identify traces of TNT or similar explosives to the 100 parts per quadrillion level, their detection rate is far more accurate and efficient. Nomadics, a Still-water, OK-based company, produces a cigar-box-sized e-nose for this purpose. Tufts University produces an optical e-nose that is designed and functions much like a mammalian nose.

Environmentalists use e-noses to analyze air. For instance, e-noses can report the chemical makeup of odors emitted by a farm’s store of manure, detect the compounds causing that odor, help minimize leaks, and determine a new diet that will decrease such odors. With their ability to detect toxic VOCs and compounds leaking from a factory’s or waste site’s storage areas, e-noses will help environmentalists force industry to change its practices. The major difficulty here remains sampling, as concentrations vary with time and place.

Caltech has used Department of Defense funding to develop a device that identifies odors in seconds. Its 32 components swell like sponges when exposed to a particular vapor, and its resistance (hence conductivity) changes accordingly. As it can detect any type of odor, doctors at the Children’s Hospital in Los Angeles are studying medical applications. Currently, it is applied to patients’ breath to help diagnose upper respiratory infections.

The major advantages of e-noses over human noses in these areas are objectivity; ability to measure odors over long real-time periods; and immunity to fatigue, infection, mental state, hazardous material, and adaptation (gradual loss of sensitivity).

How E-noses Work

E-noses have three functional components: a sample handler, a gas sensor array, and a signal processing system. Its output identifies the odorant, estimates its concentration, and relates its characteristic properties. A sensor recognizes different types and concentrations of odors through its arrays, each of which has a different sensitivity. The resulting combination provides the response pattern that enables the e-nose to identify odorants.

In a typical e-nose, a vacuum pump pulls the first air sample into the tube housing the electronic sensor arrays. The air sampling unit exposes the odorant to the sensor, after which VOCs interact with the surface and the sensor’s active material until reaching a steady state. The sensor’s response is recorded and transmitted to the signal-processing unit. When completed, a washing gas cleanses the sensor. After the reference gas is applied to the unit, the sensor is ready to measure again.

E-nose Technologies

The sensor is the e-nose’s key element, and the sensor type is its defining characteristic. There are 5 types of e-nose sensors, as follows:

Optical sensors: Optical fiber sensors work through fluorescence and chemoluminescence. The tube’s glass fibers contain a thin encoated active material in their sides and at both ends. As VOCs interact with the organic matrix’s chemical dyes, the dye’s fluorescent emission changes the spectrum. These changes then are measured and recorded for different odorous particles.

Fiber arrays with different dye mixtures can be used as sensors. These are fabricated by dipcoating (binding a plastic solution to a substrate), micro electromechanical system (MEMS), and precision machining. The main advantage is that this adjustable tool can filter out noise. Also, since many dye forms are available in biological research, sensors are cheap and easy to fabricate. But the instrumentation control systems are complex, which adds to the cost, and have a limited lifetime due to photo bleaching (the sensing process slowly consumes the fluorescent dyes).

Optical sensors are sensitive and can measure low ppb (parts per billion); however, they are still in the researach stage of development.

Spectrometry-based Sensors: This group consists of a molecular spectrum-based gas chromatography (GC), an atomic mass spectrum-based mass spectrometry (MS), and a transmitted light spectrum-based light spectrum (LS). The first two can analyze the odor’s components accurately, which is a plus. However, their use of a vapor trap to increase concentration can alter the odor’s characteristics. LS devices do not consume the sample, but do require tunable quantum-well devices. GC and MS devices are commercially available, while LS devices are only at the research stage. All spectrometry-based sensors are fabricated by MEMS and precision machining, and can measure odors to a low ppb level.

The GC tube decomposes the odorant into its molecular constituents, and MS forms a mass spectrum for each peak. The spectra then is compared to a large precompiled database of spectral peaks to classify and identify odorants.

MOSFET (Metal-oxide-silicon field-effect-transistor): The basic principle here is capacitive charge coupling. In other words, VOCs react with the catalytic metal and thereby alter the device’s electrical properties. The device’s selectivity and sensitivity can be fine-tuned by varying the metal catalyst’s thickness and composition. MOSFETs are micro-fabricated and commercially available, but can measure only parts per million. They can be manufactured by electronic interface circuits, which minimizes batch-to-batch variation. However, the gas produced by the VOC-metal reaction must penetrate the MOSFET’s gate.

Conductivity Sensors: The sensor types used here are metal oxide or conducting polymer. Both operate on the principle of conductivity, for their resistance changes as they interact with VOCs. Metal oxide sensors are common, commercially available, inexpensive, and easy to produce (they are micro-fabricated). Their sensitivity ranges from 5-500 ppm. However, they only operate at high temperatures (200°C to 400°C).

In conducting polymer sensors, VOCs bond with the polymer backbone and change the polymer’s conductivity (resistance). They are micro-fabricated together with electroplating and screen printing, are commercially available, and can measure from .1 to 100 ppm. They operate at room temperature, yet are very sensitive to humidity. Moreover, it is hard to electropolymerize the active material, which makes batch-to-batch variation inevitable. Sometimes VOCs penetrate the polymer chain, which means that the sensor must be returned to its neutral and reference state-a very time-consuming process.

Piezoelectric Sensors: These devices, which measure any change in mass, come in two varieties: quartz crystal microbalance (QCM) and surface acoustic wave (SAW) devices.

QCM sensors have a resonating disk and metal electrodes on each side. While applying the gas sample to the resonator’s surface, the polymer surface absorbs VOCs from the environment. Thus its mass increases, which increases resonance frequency. As the U.S. Navy has long used QCMs, this technology is familiar, developed, and commercially available. A QCM sensor is fabricated by screen-printing, wire bonding, and MEMS. Althoug it can measure a 1.0 Ng mass change, its MEMS fabrication and interface electronics is a major disadvantages. QCM sensors are quite linear in mass changes, their sensitivity to temperature can be adjusted, and their response to water can vary for the material used.

MEMS techniques should be handled carefully, for the surface-to-volume ratio increases drastically as dimensions approach the micrometer levels. Measurement accuracy is lost when the increasing surface-to-volume ratio begins to degrade the signal-to-noise ratio. This problem occurs in most micro-fabricated devices. SAW devices have much higher frequencies. Since 3-D MEMS processing is unnecessary, SAW devices are cheaper. As with QCM devices, many polymer coatings are available. The differential devices can be quite sensitive. However, interface electronics require more complex electronics than those of conductivity sensors for both QCM and SAW sensors. Also, as the active membrane ages, resonance frequencies can drift and so must be detected for frequency by time. SAW devices are commercially available and sensitive to mass changes at the 1.0 pg level.

Pattern Recognition

Any e-nose’s primary task is to identify an odorant and perhaps measure its concentration. After the signal processing step comes the crucial step of pattern recognition: preprocessing, feature extraction, classification, and decision-making. A database of odors must be formed for comparison purposes.

Preprocessing accounts for sensor drifts and reduces sample-to-sample variation. This can be done by normalizing sensor response ranges, manipulating sensor baselines, and compressing sensor transients.

Feature extraction involves dimensionality reduction, a crucial step for statistical data analysis, since the database’s examples usually are subject to financial constraints. The higher dimensionality caused by sensor arrays is reduced to relevant pattern-recognition information and thus extracts only significant data. As most dimensions are correlated and dependent, it is better to reduce dimensionality to a few informative axes.

Feature extraction usually is accomplished by classical principal component analysis (PGA) or linear discriminant analysis (LDA). PCA is a linear transformation that finds the maximum variance projections and the most widely used technique for feature extraction. But as PCA ignores class labels, it is not an optimal technique for odor recognition.

LDA seeks to maximize the distance between class label examples and minimize the within distance, and thus is a more appropriate approach. LDA is also a linear transformation. For instance, LDA might better discriminate subtle but crucial odor projections, whereas PCA can remove the high variance random noise in a projection.(4)

The classification stage identifies odors. Classical classification techniques are KNN (k nearest neighbors), Bayesian classifiers, and ANN (artificial neural networks]. KNN with, say, 5 nearest points will find the 5 closest matches from the precompiled database. The closest match will be assigned as the tested material’s odorant class.

Bayesian classifiers first assign a posterior probability to the classes in the lower dimension and then pick the class that maximizes the predetermined probability distribution. ANN is closer to biological odor recognition. After being trained by the odor database, it is exposed to the unknown odorant in order to recognize the largest applicable response odorant class. The classifier estimates the class and places a confidence level on it.

In decision-making, risks and application-specific knowledge are considered in order to modify the classification. All decisions are reported-even a nonmatch.

Conclusion

As this article indicates, we can expect great progress in this area. And with each step forward, science and technology will continue to point toward the Greatest Artist’s most subtle designs and allow us to appreciate them better.

Footnotes

1. There are over 100 million receptor proteins of about 1,000 different types.
2. A human olfactory bulb contains approximately 2,000 glomeruli.
3. The terms “electronic nose” and “e-nose” are incorrect, for these devices cannot be considered “real” noses. The correct terminology should be “electronic arrays for chemical sensory and identification.” However, “e-nose” has gained wide acceptance in the literature since it first appeared during a 1991 NATO workshop in Reykjavik, Iceland.
4. Such nonlinear transformations as Sammon nonlinear maps and Kohonen self organizing maps also are used in feature extraction. These preserve the distance between pairs of examples when reducing dimensionality to 2 or 3.

References

• Baltes, Henry, Dirk Lange, and Andreas Koll. “The Electronic Nose in Lilliput.” IEEE Spectrum (Sept. 1998): 35-38.
• Barinaga, Marcia. “Salmon Follow Watery Odor Home.” Science 286 (22 Oct. 1999): 705-6.
• http://csmt.jpl.nasa.gov/enose.html.
• http://faculty.washington.edu/chudler/nosek.html.
• Malakoff, David. “Following the Scent of Avian Olfaction.” Science 286 (22 Oct. 1999): 704-5.
• Mamberts, Peter. “Seven-Transmembrane Proteins as Odorant and Chemosensory Receptors.” Science 286 (22 Oct. 1999): 707-10.
• Perkins, Sid. “Eau, Brother! Electronic Noses Provide a New Sense of the Future.” Science News 157 (19 Feb. 2000): 125-27.
• Schiffmann, Susan, and H. Troy Nagle. “The How and Why of Electronic Noses.” IEEE Spectrum (Sept. 1998): 22-32.
• Stern, Peter and Jean Marx. “Making Sense of Scents.” Science 286 (22 Oct. 1999): 703.
• Wolfgang, Gopel, and Tilo Weiss. “Design for Smelling.” IEEE Spectrum (Sept. 1998): 32-34.
• www.planetee.com/planetee/servlet/DisplayDocument?ArticleID=6899.
• www.sfn.org/briefings/smell.html.