The Potato Operation is a collaborative, multidisciplinary project in the domain of destructive testing of agricultural products. It aims at automatizing pulp sampling of potatoes in order to detect possible viral diseases. Such viruses can decrease fields productivity by a factor of up to ten. A machine, composed of three conveyor belts, a vision system, a robotic arm and controlled by a PC has been built. Potatoes are brought one by one from a bulk to the vision system, where they are seized by a rotating holding device. The sprouts, where the viral activity is maximum, are then detected by an active vision process operating on multiple views. The 3D coordinates of the sampling point are communicated to the robot arm holding a drill. Some flesh is then sampled by the drill, then deposited into an Elisa plate. After sampling, the robot arm washes the drill in order to prevent any contamination. The PC computer simultaneously controls these processes, the conveying of the potatoes, the vision algorithms and the sampling procedure. The master process, that is the vision procedure, makes use of three methods to achieve the sprouts detection. A profile analysis first locates the sprouts as protuberances. Two frontal analyses, respectively based on fluorescence and local variance, confirm the previous detection and provide the 3D coordinate of the sampling zone. The other two processes work by interruption of the master process.

A system for the simultaneous acquisition of multiple images using laser scanning techniques is reported. It is intended for inspection tasks that require a variety of different image qualities to be highlighted. The particular application this system has been designed for is the grading of sheep pelts prior to sale for processing into leather. The primary aim of the application is the presentation of consistent and relevant information to an operator, allowing better grading decisions to be made. To achieve this, three different types of images are generated and presented. These are; reflection, transmission and fluorescence, all being acquired simultaneously using a scanned laser spot. Techniques used to collect, concentrate and detect the laser and fluorescent light are outlined and images presented. Ultimately, it is hoped that the machine will automatically detect defects and classify pelts. To this end high speed computing hardware is incorporated into the system design.

Quick identification of microorganisms is an essential work for researchers and industries. Motion characteristics of microorganisms carry important information for distinguishing them. It is obvious that developing a test technique to identify individual microorganism under a living condition which is closer to the natural state will have theoretical and practical significants. In this paper an image tracking system for approach on the living microorganisms is described. The system consists basically of a set of optical image device, a CCD photodetector and an image analysis system. A compact configuration of industrial microorganism motion image tracking system was proposed for applying on-line inspection.

A PC-based machine vision system providing rapid measurement of bare-root tree seedling morphological features has been designed. The system uses backlighting and a 2048-pixel line- scan camera to acquire images with transverse resolutions as high as 0.05 mm for precise measurement of stem diameter. Individual seedlings are manually loaded on a conveyor belt and inspected by the vision system in less than 0.25 seconds. Designed for quality control and morphological data acquisition by nursery personnel, the system provides a user-friendly, menu-driven graphical interface. The system automatically locates the seedling root collar and measures stem diameter, shoot height, sturdiness ratio, root mass length, projected shoot and root area, shoot-root area ratio, and percent fine roots. Sample statistics are computed for each measured feature. Measurements for each seedling may be stored for later analysis. Feature measurements may be compared with multi-class quality criteria to determine sample quality or to perform multi-class sorting. Statistical summary and classification reports may be printed to facilitate the communication of quality concerns with grading personnel. Tests were conducted at a commercial forest nursery to evaluate measurement precision. Four quality control personnel measured root collar diameter, stem height, and root mass length on each of 200 conifer seedlings. The same seedlings were inspected four times by the machine vision system. Machine stem diameter measurement precision was four times greater than that of manual measurements. Machine and manual measurements had comparable precision for shoot height and root mass length.

A PC-based machine vision system was used at a forest nursery for two months to make quality control measurements of bareroot pine seedlings. In tests conducted during the lifting season, there was close agreement between machine vision and manual measurement distribution results for seedling samples for both root collar diameter and tap root length. During a second set of tests conducted after adding a bud tip height measurement routine, measurement distribution results for seedling samples were in close agreement for root collar diameter, tap root length, and bud tip height. Machine vision measurements of root collar diameter and tap root length also correlated well with manual measurements on a seedling-to- seedling basis for the second test. With the machine vision system, seedling samples could be measured by one person in approximately the same amount of time that it took two people to measure them manually.

This paper describes a system for real-time inspection of 2D surfaces. It was initially planned as system for classification of wooden surfaces, but was successfully used also in the context of other inspection tasks like metallic surface inspection and leather inspection. The system has two major modules. One is a 2D object segmentation and recognition part, where key elements of the underlying elements have been published before. This includes hierarchical processing of the incoming gray-level images leading to a symbolic description of the surface; syntactic segmentation; and the decision network methodology used. Beyond these features, a new track has been added, which is entirely devoted to texture classification in real-time. This two-way analysis of wooden surfaces was first implemented on a heterogeneous architecture containing Zoran vector processors and Transputers (all commercially available). The current version uses only TMS32C40 processors. The system has been successfully implemented in a production plant in Austria. We describe major elements of the system and the underlying algorithms.

Somatic embryogenesis is a process considered very promising in mass regeneration of plant materials. Quality evaluation of embryos' viability is deemed necessary during the process. Machine vision techniques have been or are being developed for quality sorting of embryos immediately before germination. In this project, a robotic workcell has been conceptually designed for quality sorting of somatic embryos employing machine vision system. The configuration of the workcell has been modeled using a 3D graphic modeling software package. The workcell component requirements, layout, and materials flow have been investigated for its workability. A numerical model has been developed to simulate the productivity of a workable layout. Some stochastic factors affecting the workcell productivity were considered in the numerical model. Engineering economic analysis was performed on the workcell for evaluating its cost-effectiveness.

Wood surface roughness is one of the important indexes of manufactured wood products. This paper presents an attempt to develop a new method to evaluate manufactured wood surface roughness through the utilization of imaging processing and pattern recognition techniques. In this paper a collimated plane of light or a laser is directed onto the inspected wood surface at a sharp angle of incidence. An optics system that consists of lens focuses the image of the surface onto the objective of a CCD camera, the CCD camera captures the image of the surface and using a CA6300 board digitizes the image. The digitized image is transmitted into a microcomputer. Through the use of the methodology presented in this paper, the computer filters the noise and wood anatomical grain and gives an evaluation of the nature of the manufactured wood surface. The preliminary results indicated that the method has the advantages of non-contact, 3D, high-speed. This method can be used in classification and in- time measurement of manufactured wood products.

Evaluation of bread crumb texture plays an important role in judging bread quality. This paper discusses the application of image analysis methods to the objective measurement of the visual texture of bread crumb. The application of Fast Fourier Transform and mathematical morphology methods have been discussed by the authors in their previous work, and a commercial bread texture measurement system has been developed. Based on the nature of bread crumb texture, we compare the advantages and disadvantages of the two methods, and a third method based on features derived directly from statistics of edge density in local windows of the bread image. The analysis of various methods and experimental results provides an insight into the characteristics of the bread texture image and interconnection between texture measurement algorithms. The usefulness of the application of general stochastic process modelling of texture is thus revealed; it leads to more reliable and accurate evaluation of bread crumb texture. During the development of these methods, we also gained useful insights into how subjective judges form opinions about bread visual texture. These are discussed here.

The distribution and migration of moisture in wheat kernels during storage and milling are important for controlling wheat quality and milling performance. A non-destructive microscopic magnetic resonance imaging (MRI) technique was applied to measure the 3D distribution of moisture in individual wheat kernels. Soft white winter wheat kernels at approximately 12% moisture content were measured. A 3D Fourier transform pulse sequence with a short echo time (TE) was used to acquire proton images. The image data were zero- filled to improve spatial resolution. The spatial resolution of the images was 78 micrometers X 62 micrometers X 62 micrometers . The 3D proton density images were related to the 3D moisture distribution in the wheat kernels. For wheat kernels equilibrated to a water activity of 0.53 a_w, the moisture distribution in the starchy endosperm of the wheat was fairly uniform in the three principal (orthogonal) directions. To quantify moisture distribution and movement in wheat kernels during the moisture-tempering process, the microscopic MRI with a shorter TE 3D pulse sequence will be used in future research. The results of the measurements will be used to characterize the quantity and nature of kernel moisture and relate these to kernel physiology, physical-chemical properties, and milling performance.

A machine vision system was developed to separate stained pistachio nuts, which comprise of about 5% of the California crop, from unstained nuts. The system may be used to reduce labor involved with manual grading or to remove aflatoxin contaminated product from low grade process streams. The system was tested on two different pistachio process streams: the bi- chromatic color sorter reject stream and the small nut shelling stock stream. The system had a minimum overall error rate of 14% for the bi-chromatic sorter reject stream and 15% for the small shelling stock stream.

Pattern recognition techniques have some attraction for the automatic identification of seeds as they are fast, non-destructive and easily applied. In this paper, the performance of quadratic discriminant functions and one form of artificial neural network are compared for the task of identifying Australian wheat varieties. This is a complex problem as the kernels are very similar in appearance, and factors other than variety significantly influence shape. It is shown both approaches have some similarity in performance, but discriminant functions provide a superior result and are more easily applied. There is, though, some opportunity for further refinement of the artificial neural network.

Because assessment of Durum wheat semolina purity by standard ash-test has been widely criticized, we attempted to characterize products of a semolina mill by spectrofluorometry and fluorescence imaging. A collection of milled wheat products ranging from very pure semolina to brans were chosen for this study. Multidimensional statistical analyses (Principal component analyses) were applied to the spectral and image data. Maps showing a classification of the products according to purity were obtained without biochemical calibration. Principal component regression was applied to the data in order to test the relationship of aleurone fluorescence to ash content. Both spectrofluorometry and fluorescence imaging gave similar results with good determination coefficients (r² equals 0.97 and 0.92) for the study of a single wheat variety. Products obtained from different wheat varieties were more difficult to compare.

A methodological approach of pattern recognition is proposed to make on-line selection of cereal products as they are poured down through the air. It addresses the characterization of product populations according to different pre-defined quality classes using image analysis. In order to take into account the complexity of this problem, we propose to use both global (related to the whole image) and individual (related to each particle) variables. The aim of the data processing module is to reduce the dimension of the variable space without losing information, and then to select the most pertinent components to train the decision system. The first stage of this system, called generalist one, has to give an ambiguous response, that means to select a subset of possible output classes. The second, or specialist, which is trained to distinguish only some subjects of classes delivered by the generalist one, gives the decision. The method has been applied in the framework of milling products classification. Three quality classes have been defined, they correspond to the rolls gap (0.30, 0.40, 0.50 mm.) of the first break rolls of a semolina pilot mill. In these conditions, the classification accuracy rate achieved by the system is higher than 80%.

An integrated optic biosensor for detecting foodborne pathogens is described. The sensor is based on a planar waveguide operating in an interferometric mode. The device functions by detecting the direct binding of an antigen molecule to a functionalized waveguide surface. It is capable of detecting biomolecules at subnanogram/milliliter concentrations and has been used to detect proteins specific to Salmonella.

Theis method facilitates and makes convenient and portable X..ray imaging ofsmall biological objects (grains, leaves, insects, etc.) almost instantly and without any wet developing processes. The key technology that makes that possible is photothermoplastic carrier (film) that is sensitive to the X-rays. Small biological object is placed into an imaging camera, exposed to an X-ray source and the image is registered on the photothermoplastic film. The development process is very simple heating the film to approx.70degrees Celsius fully develops the image, which remain stable after cooling. The camera is ready to the next job. The whole device could be made quite portable and can be used in field research. Our first experimental portable system weigh about 40 lb.

The paper presents a granule appearance recognition system in the plastic production process. In the industry, granule product shape is one of the important quality index. Abnormal granule shape products are as waste products some times. But in the automatic process, it is too difficult to test the granule shape one by one, when the product quantity is great. The automatic recognition system we proposed is designed for distinguishing the normal granule or abnormal granule shape of plastic products in the process line and giving the percentages of them. The methods (8 chain code, automatic threshold selection...) we used realize the rapid granule shape test, therefore the good quality products are obtained.

An integrated system which consisted of a visible/near-infrared spectroscopic subsystem and an intensified multispectral imaging subsystem was tested for its accuracy in separating abnormal (unwholesome) from normal poultry carcasses. The spectroscopic subsystem measured reflectance spectra of the poultry carcasses at wavelengths from 471 to 965 nm. For the multispectral imaging subsystem, the gray-level intensity of whole carcasses was measured using six different optical filters of 542, 571, 641, 700, 720, and 847 nm wavelengths. The preliminary results showed that, with the integrated system, there were no abnormal carcasses being misclassified as normal carcasses. When individual subsystem was used for classification, the error of the spectroscopic subsystem was 2.6% and that of the multispectral imaging subsystem was 3.9%. Thus, the integrated system could be used for separating carcasses into normal and abnormal streams. With perfect selection of normal carcasses in the normal carcass stream, the inspector needs only inspect the abnormal carcass stream. Thus, the through-put of carcasses of the processing line per inspector could be greatly increased.

Results of research to develop a reference method for calibrating standard dot image tiles and cotton trash boxes are reported. The tiles and trash boxes are produced by the Agricultural Marketing Service for use in standardizing HVI Video Trashmeters. The reference method involves use of a highly sensitive image analysis system (the Quantimet 970) to measure the number and percent area fraction of particle images produced on the replica tiles or of real trash particles placed on the surface of cotton incased in plastic boxes. Calibration data for a set of tiles and boxes is presented. The effects of magnification (system resolution) and detection threshold are investigated as related to measurement accuracy.

A unified method for detecting all types of textural faults on a carpet using machine vision is presented. The Gaussian Markov Random Field (GMRF) model is used for the modelling of the textural surface of carpet. An experimental device using a line-scan camera and an IBM personal computer has been set up simulating on-line inspection of woven carpets to detect various types of fault arising in the production process. Measures for detecting faults are derived from the GMRF model based on sufficient statistics. This measure is very effective in detecting textural differences. Detection of unlevel, linear and other types of faults is discussed. In combination with our previous linear faults detection method, we have the confidence to be able to detect all types of textural faults on a plain carpet in an efficient way. With some additional techniques, this method can also be used for the detection of faults in colored pattern carpets.

Insect infestation increases the probability of aflatoxin contamination in pistachio nuts. A non- destructive test is currently not available to determine the insect content of pistachio nuts. This paper uses film X-ray images of various types of pistachio nuts to assess the possibility of machine recognition of insect infested nuts. Histogram parameters of four derived images are used in discriminant functions to select insect infested nuts from specific processing streams.

French fries are one of the frozen foods with rising demands in domestic and international markets. Color is one of the critical attributes for quality evaluation of french fries. This study discusses the development of a color computer vision system and the integration of neural network technology for objective color evaluation and classification of french fries. The classification accuracy of a prototype back-propagation network developed for this purpose was found to be 96%.

Digitized images from black and white video cameras are being used to measure the area and numbers of leaf particles in cotton after lint cleaning. The method is now used to provide trash grades for 16 - 18 million bales of cotton prepared for market each year. Small samples are compressed against a glass window and illuminated with two small incandescent lamps for imaging. Leaf area readings are automatically adjusted for differences in lint greyness. The accuracy of this method compares well with gravimetric measurements of non-lint content.

The High Resolution Picture Transmission (HRPT) and Automatic Picture Transmission (APT) data from the National Oceanic and Atmospheric Administration weather satellite Advanced Very High Resolution Radiometers were the primary sources of satellite data used in this study. Several models were developed to convert the gray-level digital numbers of the satellite imagery into temperature values. The satellite-derived temperatures, which were stored in a geographic information system database, were well correlated with ground-truth temperature measurements. The results of this study indicated that the APT data is useful for lake temperature estimation as compared with the HRPT data. In other words, both HRPT and APT data are capable of being converted into lake surface temperature with acceptable accuracy.

Crop surface temperature under a radiant heated greenhouse was measured using a portable infrared thermometer. Plants were arranged so that each plant occupied a grid cell of 30 cm X 30 cm (1 ft X 1 ft). Data collected were analyzed for their spatial distribution. Geostatistical software was used to characterize the spatial variability of the plant surface temperature. The shape of the empirical semi-variogram suggested that a spherical model was best fitted to the empirical semi-variogram. This model indicated that the nugget effect was estimated at 1.2, the sill at 3.3 and the range at 1.65 m. This model was used in block kriging to estimate plant surface temperature for unsampled locations.

Spectral and dynamic morphological features were investigated for plant health monitoring using machine vision techniques. The plants were stressed by withholding all nutrient salts. The spectral reflectance of healthy and stressed lettuce leaves (Latuca sativa cv. `Ostinata') was measured to determine at which wavelength(s) a stressed condition would be apparent. The measured wavebands were between 400 and 1000 nm. A reference waveband was utilized to account for photometric variables such as lighting and surface geometry differences during image acquisition. The expansion of the top projected leaf area (TPLA) was found to be an effective feature to identify stressed plants. The nutrient stressed plant was identifiable within two days after nutrients were withheld from a healthy plant. This was determined by a clearly measurable reduction in TPLA expansion.

The method for on board calibration of the space imaging instruments using a diffusing target illuminated by the sun radiation is gaining acceptance. This raises a problem of measurements of the spectral and angular response of the target reflectivity in illumination conditions similar to the actual one on-board. This work summarizes the progress in the development and testing of the target for on-board radiometric calibration of the space video-spectrometers in stereo- spectral imaging system ARGUS. The spectral and angular response of the target reflectivity and the experimental procedure used for getting these characteristics are described. The target manufacturing technique is described. This target was tested for the spectral properties from 0.3 micrometers to 5.2 micrometers at angle of illumination -70 degree(s) (relative to the normal of the target) and angle of viewing +20 degree(s). The angular response was measured with three incident angles 70 degree(s), 75 degree(s) and 80 degree(s) (relative to the normal) for angles of viewing varying between 0 degree(s) and 60 degree(s) (relative to the normal). Brief description of the setup for measurements of the spectral and angular response of the target reflectivity are presented. The values of angular and spectral response of the reflectivity have been determined with the accuracy of 2 - 5% depending on the wavelength.

Optical plant sensors constructed from red and near-infrared (NIR) filtered photodetector pairs were used in conjunction with the normalized difference index (NDI) to detect plants. Plants must occupy a minimum of 8% of the photodetector pair field-of-view (FOV) to be detected. Thus, knowing the size and location of the FOV is crucial. Since the NDI requires red and NIR reflectance measurements from coincident areas, it is equally important to know the coincident area of a red-NIR detector pair and its location. Reflectance measurements taken every 1 cm from a 60 cm X 60 cm surface can be graphically viewed to determine the size and location of the FOV of a plant sensor. The surface has low reflectance and contains a 20 cm X 20 cm highly reflective checkerboard pattern in the center. From individual FOVs of red-NIR pairs, the coincident area can be found.

Micropropagation of plants is done by cutting juvenile plants and placing them into special container-boxes with nutrient-solution where the pieces can grow up and be cut again several times. To produce high amounts of biomass it is necessary to do plant micropropagation by a robotic syshoot. In this paper we describe parts of the vision syshoot that recognizes plants and their particular cutting points. Therefore, it is necessary to extract elements of the plants and relations between these elements (for example root, shoot, leaf). Different species vary in their morphological appearance, variation is also immanent in plants of the same species. Therefore, we introduce several morphological classes of plants from that we expect same recognition methods. As a result of our work we present rules which help users to create specific algorithms for object recognition of plant species.

This paper describes the measurement and range image processing of soil/stubble. The purpose of the data processing is to identify the data points corresponding to standing straw stubble from the soil surface in order that the erosion hazard of the soil can be studied. The 3D measurement system consists of a camera and a projector. A series of light stripes is projected onto the soil. By observing sample points on the camera image of the stripes on the soil surface, the relative positions of the corresponding points on the soil surface can be calculated. The system measures up to 8000 sample points in about half a minute. Intuitively, if the soil surface is aligned to a horizontal surface, the samples on the straws correspond to points with high altitudes. Thus, the first step in the processing is to construct the underlying soil surface. By comparing the difference in the height of the samples and the constructed soil surface, the samples corresponding to the straws can be identified. Discrete cosine transform low-pass filtering and morphological opening are examined for the soil surface construction. It is shown by simulation that the morphological opening provides much better results. The algorithm with morphological opening is also tested with measurement data. Preliminary results show that the algorithm provides a workable way to segment soil surface data and stubble data from the measurement.

Development of a machine vision device to automatically identify different weed species within a field is needed to design a successful spatially variable herbicide applicator. This study was conducted to develop a computer vision algorithm that can successfully identify a broadleaf weed (velvetleaf, Abutilon theophrasti), a grassy weed (wild proso millet, Panicum miliacem), and corn (Zea mays, L.). Digital images were collected in laboratory and field conditions for all three plant species. Image analysis techniques were used to analyze the possibility of using a combination of size and shape features to produce a classification scheme. Two separate approaches were used to classify the velvetleaf from the wild proso millet and corn, and the wild proso millet from the corn. The first and second invariant central moment of inertia measurements along with plant perimeter were used to separate the velvetleaf from the monocot species. Due to the similar shapes of wild proso millet and corn, we were unable to classify the two species by only using size and shape features. Consequently, a two step approach was utilized. This involved using projected perimeter to determine the age (number of days after emergence) of the plant. By knowing the possible age of the plant, the wild proso millet and corn were classified using a combination of length and circularity. Future research will involve the evaluation of several other image features to determine the best classification scheme. Further data will also be collected from a library of laboratory and field images in order to increase the confidence interval of the classification scheme.

Seed spacing uniformity is crucial to crops such as sugarbeets. Planter manufacturers, dealers, and sugarbeet growers are interested in how uniformly planters place seeds into the soil. The current laboratory method for testing planter accuracy uses a stationary planter unit to place seeds onto a grease belt, allowing visualization of the seed spacings. This method, however, is very difficult and time consuming. Automation of the test using photonics is desirable. A rectangular photogate block was designed with light emitting diodes (LEDs) opposite photo- transistors. Twenty-four pairs of LEDs and photo-transistors were connected to a digital I/O board on a personal computer to detect space and time between seeds. During laboratory testing of a planter unit, the photogate was positioned beneath the seed drop tube to simulate the bottom of a furrow. A program was developed to poll the twenty four digital channels and three counters so no seeds could go through the photogate without being detected. Planter ground speed was monitored through a magnetic switch. Histograms of seed spacing and relative location of seed passing through the photogate were provided for each test. The outputs conform with the requirements of International Organizations for Standardization 7256/1 - 1984.

An integrated optic sensor for monitoring NH₃ volatization as related to agricultural fertilizer applications is described. The sensor is capable of monitoring NH₃ levels over a range from less than 100 parts per billion by volume to levels approaching 1000 parts per million by volume. The sensor is based on a planar waveguide operating in an interferometric mode. The device functions by monitoring a refractive index change resulting from a reversible chemical reaction occurring on the waveguide surface.

This paper presents a method of estimating the lean and fat composition in pork ham from cross-sectional area measurements using image processing technology. The relationship between the quantity of ham lean and fat mass with the ham lean and fat areas was studied. The prediction equations for pork ham composition based on the ham cross-sectional area measurements were developed. The results show that ham lean weight was related to the ham lean area (r equals .75, P < .0001) while ham fat weight was related tot the ham fat area (r equals .79, P equals .0001). Ham lean weight was highly related to the product of ham total weight times percentage ham lean area (r equals .96, P < .0001). Ham fat weight was highly related to the product of ham total weight times percentage ham fat area (r equals .88, P < .0001). The best combination of independent variables for estimating ham lean weight was trimmed wholesale ham weight and percentage ham fat area with a coefficient of determination of 92%. The best combination of independent variables for estimating ham fat weight was trimmed wholesale ham weight and percentage ham fat area with a coefficient of determination of 78%. Prediction equations with either two or three independent variables did not significantly increase the accuracy of prediction. The results of this study indicate that the weight of ham lean and fat could be predicted from ham cross-sectional area measurements using image analysis in combination with wholesale ham weight.

A test has been run to study the efficacy of detecting particulate contaminants in processed meat samples by manual observation of line-scanned x-ray images. Six hundred processed product samples arriving over a 3 month period at a national USDA-FSIS laboratory were scanned at 230 cm²sec with 0.5 X 0.5 mm resolution, using 50 KV, 13 ma excitation, with digital interfacing and image correction. Images were inspected off-line, using interactive image enhancement. Forty percent of the samples were spiked, blind to the analyst, in order to establish the manual recognition rate as a function of sample thickness [1 - 10 cm] and texture of the x-ray image [smooth/textured], as well as spike composition [wood/bone/glass], size [1 - 4 mm] and shape [splinter/round]. The results have been analyzed using maximum likelihood logistic regression. In meat packages less than 6 cm thick, 2 mm bone chips are easily recognized, 1 mm glass splinters with some difficulty, while wood is generally missed even at 4 mm. Operational feasibility in a time-constrained setting has bee confirmed. One half percent of the samples arriving from the field contained bone slivers > 1 cm long, one half percent contained metallic material, while 4% contained particulates exceeding 3.2 mm in size. All of the latter appeared to be bone fragments.

We report on a new class of instrument for imaging spectral analysis, the tunable liquid crystal filter (LCTF). The LCTF is an optical filter, similar to an interference filter, whose center wavelength is electronically tunable with no moving parts, in a few milliseconds, across hundreds of nanometers. The filter is a polarization interference filter based on the Lyot design, using the electro-optic action of liquid crystals to tune the passband. Imaging quality is near the diffraction limit and there is no image shift as the filter is tuned. Bandwidths ranging from a 0.25 nm to 60 nm have been achieved, for use in high-resolution sequential RGB imaging, microscopy of multiply-tagged fluorescent samples, bathymetry, and remote sensing. LCTFs are presently being applied to agricultural quality control measurements.

Near infrared reflectance (NIR) images of bruised `Delicious' applies were converted to images of texture properties. Bruises of two sizes (11 mm and 26 mm diameter) and two ages (1 d and 90 d) were examined. Seven texture properties (variance, entropy, product moment, difference entropy, inverse difference, difference variance, and sum variance) were computed from a cooccurrence matrix. Window size and neighborhood distance for the cooccurrence matrix were set to optimize the texture contrast between bruised and unbruised tissue. The window position was incrementally scanned over the entire apple image creating a new image of texture values. Four neighborhood directions (0 degree(s), 90 degree(s), 45 degree(s), 135 degree(s)) were considered. Sum variance was the only texture property that showed improved contrast of the bruised/unbruised areas relative to the original NIR image. All other texture properties produced images that highlighted the edge of the bruise. The variance property produced images with the best defined bruise edges irregardless of bruise size or age. Variance and sum variance show promise as additional features to the grey tone image for discriminating apple bruises.

Light transmittance in the 450 to 1050 nanometer (nm) region was evaluated as a nondestructive technique for identifying apples with internal browning. Shorter wavelengths of light (< 800 nm) were attenuated more than longer wavelengths (> 800 nm). A transmission difference between 720 and 810 nm was used to segregate apples with internal browning from good apples. Only 7.4% of the apples were misclassified in a training set. When applied to a larger validation set, 8.0% of the apples with internal browning were misclassified. For both sets, the only apples misclassified were those with very slight browning that was very difficult to detect visually were misclassified, but none of the apples with slight to severe browning was misclassified. For nondefective apples, 6.1% were identified as having internal browning, because bruises and internal browning had the same effect on the spectral composition.

A spectrophotometer system is being updated with a new computer and control circuitry to measure agricultural products in the UV through NIR regions of the electromagnetic spectrum. This single beam instrument, along with its computer, will have wavelength repeatability near 0.005 nm. This performance is required for proper system response compensation in a single beam system. Analog electronics are kept to a minimum by early conversion of the signal with a 10-microsecond(s) 16-bit A/D converter. A fast response real-time computer is required to service the A/D, wavelength drive, and optical chopper. The paper reports on the development and selection of computer interfaces, data collection techniques, and performance characteristics of this laboratory spectrophotometer system. The system is regularly used as the development tool in investigating new measurement techniques on agricultural products, and also to evaluate optical filters and other spectrophotometric systems.

Methods of nondestructively examining Granex type sweet onions are needed to insure that only good quality onions are shipped at harvest and to avoid putting infected onions in controlled atmosphere (CA) storage where they occupy valuable space and can ruin surrounding onions. A Toshiba TCT 20Ax tomographic scanner operated in the line scan mode and an incandescent light box were used to evaluate the potential for detecting infected onions nondestructively. A study (CA storage study) involving 200 onions, 100 harvested early and 100 harvested late, one half destructively inspected before the remaining half were placed into CA storage was initiated May 1994. All onions were line scanned and scored with the light box before CA storage and those in CA storage were line (will be) scanned and optical scored on retrieval from the storage. An additional study (Disease storage study) involving 40 onions, late harvest, stored at 25C, 60% rh for three weeks with line scanning as above on a weekly interval. After the third week these fruit were assayed for visual damage and for decay organisms. Results from the incandescent light box scoring were not encouraging. From both studies the number of defects, average defect size and the difference image intensity as determined from line scanning were the major contributing parameters to a discriminant analyses model predicting about 70% or better accuracy.

Manufacturing processes that utilize natural products as raw materials for forming their deliverables face additional challenges in the areas of quality control and inspection. This comes about from the natural variability that occurs in the products. Systems to automate this activity have been difficult to design and implement from the view of algorithm complexity which impacts the computational requirements for real-time execution. This paper will describe a technique for recognizing global or systematic defects on poultry carcasses. A method for implementing the technique that is capable of executing at a rate of about 180 birds per minute is described.

The prediction of intramuscular fat (or marbling) of live beef animals using serially scanned ultrasound images was studied in this paper. Image analysis, both in gray scale intensity domain and in frequency spectrum domain, were used to extract image features of tissue characters to get useful parameters for prediction models. One, 2 and 3 order multi-variable prediction models were developed from randomly selected data sets and tested using the remained data sets. The comparisons of prediction results between using serially scanned images and only final scanned ones showed good improvement of prediction accuracy. The correlation of predicted percent fat and actual percent fat increase from .68 to .80 and from .72 to .76 separately for two groups of data, the R squares increase from .65 to .68 and from .68 to .72, and the root of mean square errors decrease from 1.70 to 1.52 and from 1.22 to 1.12 separately. This study indicates that serially obtained ultrasound images from live beef animals have good potential for improving the prediction accuracy of percent fat.

Conventional soil maps represent the distribution of soil attributes across landscapes but with less precision than is needed to obtain the full economic and environmental benefits of site- specific crop management. This study quantifies the spatial variability of three agronomically significant soil attributes: (1) thickness of mollic epipedon, (2) organic matter content (OM), and (3) pH as related to soil survey map units, spectral data, and terrain attributes for a 20 ha field in Montana. Analysis of Order 1 (1:7920-scale) Soil Survey map units indicates substantial variation in all three soil attributes. There was some evidence that similar attribute values were clustered in the field (0.40 - 0.46 Moran's Coefficients). Two spectral band ratios explained 64% of the variation in OM across the field. GPS/GIS-derived wetness index, sediment transport index, elevation, and slope gradient explained 48% of OM variation. Wetness index, slope gradient, and plan curvature combined to explain 48% of the variation in mollic epipedon thickness. Elevation and wetness index explained just 13% of pH variation. Two spectral band ratios, specific catchment area, and wetness index combined to explain 70% of the variation in OM at 66 sampling sites. Four contour map representations of OM illustrate the sensitivity of the final maps to variations in input data and interpolation method.

The utilization of secondary energy resources in heating networks of food-producing enterprises resolves to decide an important task: to economize the utilization of fuel and energy resources.

Imaging spectrometers have recently moved out of the spaceflight environment, in which they were developed, to a host of other applications. Some of these new uses include the graphics and printing industry, process control, bio-medicine, clinical diagnostics and agriculture. For any of these applications, new approaches are necessary to design compact, portable instruments that can be easily and reliably calibrated. This paper presents one such implementation of an imaging spectrometer suitable for field use.

A fully integrated instrument for cereal grain quality assessment is presented. Color video images of grains fed onto a belt are digitized. These images are then segmented into kernel entities, which are subject to the analysis. The number of degrees of freedom for each such object is decreased to a suitable level for Artificial Neural Network (ANN) processing. Feed- forward ANN's with one hidden layer are trained with respect to desired features such as purity and flour yield. The resulting performance is compatible with that of manual human ocular inspection and alternative measuring methods. A statistical analysis of training and test set population densities is used to estimate the prediction reliabilities and to set appropriate alarm levels. The instrument containing feeder belts, balance and CCD video camera is physically separated from the 90 MHz Pentium PC computer which is used to perform the segmentation, ANN analysis and for controlling the instrument under the Unix operating system. A user-friendly graphical user interface is used to operate the instrument. The processing time for a 50 g grain sample is approximately 2 - 3 minutes.

Multispectral image analysis was used to identify broken and sound kernels in bulk samples of corn. Images in a 512 X 512 pixel format were acquired with 50 nm bandpass filters in the visual and near infrared regions of the spectrum. Samples of broken and sound corn kernels were assessed. A search for pixels which represented endosperm and sound tissue of the kernels was done by relating the gray values from different bandwidth images at the same topological location. Data analysis was done using means of 4 X 4 arrays of normalized pixel values and derived features to create a pattern for sample recognition. The most effective bandwidths for identification of endosperm tissue was determined followed by a search of pixel coordinates to identify endosperm areas. Binarization of the endosperm areas, reflective spots and shade between kernels was done as a preprocessing step. Samples were then classified by evaluation of 256 X 256 pixel subimages of each sample. A 100% correct recognition rate of the broken and sound corn classes was achieved.