With the current trend of integrating machine vision systems in industrial manufacturing and inspection applications comes the issue of camera and illumination stabilization. Unless each application is built around a particular camera and highly controlled lighting environment, the interchangeability of cameras of fluctuations in lighting become a problem as each camera usually has a different response. An empirical approach is proposed where color tile data is acquired using the camera of interest, and a mapping is developed to some predetermined reference image using neural networks. A similar analytical approach based on a rough analysis of the imaging systems is also considered for deriving a mapping between cameras. Once a mapping has been determined, all data from one camera is mapped to correspond to the images of the other prior to performing any processing on the data. Instead of writing separate image processing algorithms for the particular image data being received, the image data is adjusted based on each particular camera and lighting situation. All that is required when swapping cameras is the new mapping for the camera being inserted. The image processing algorithms can remain the same as the input data has been adjusted appropriately. The results of utilizing this technique are presented for an inspection application.

This paper describes a proposed electro-optical implementation of the lateral inhibition function and its simulation. Lateral inhibition is the most important adaptation property in sensory systems and with this mechanism, weak inputs to neurons are suppressed (inhibited) by strong inputs to neighboring neurons. The purpose of lateral inhibition is to enhance and emphasize boundaries. Two types of lateral inhibition filters were studied, both symmetric and antisymmetric. The antisymmetric are direction dependent and thus, four different directions were chosen. Since the electro-optic implementation relies on filtering input images in the Fourier plane, the inhibition filters used in the simulation are phase-only. The simulation takes into consideration performance of specific optical hardware components and was tested on signals that have very sharp directional edges in addition to a facial image with many contour line edges.

The research effort reported in this paper focuses on the evaluation of different input codings influencing the performance of a back-propagation neural network for classification of remotely sensed images. The clustering capability, which can be visualized through the Euclidean distance graph, is introduced as a tool to predict the credibility of the input coding. An investigation was also conducted to study the use of weight function to improve the clustering capability of the binary-coded-decimal input coding, a widely used coding approach in remote sensing area. Results obtained indicate that the classification performance of the neural network classifier is closely related to the clustering capability of the input codings. In order to fully exploit the generalization property of neural network, the clustering property of the classes must be maintained during the input coding process.

In Switzerland, aerial photos are indispensable tools for research into ecosystems and their management. Every six years since 1950, the whole of Switzerland has been systematically surveyed by aerial photos. In the forestry field, these documents not only provide invaluable information but also give support to field activities such as the drawing up of tree population maps, intervention planning, precise positioning of the upper forest limit, evaluation of forest damage and rates of tree growth. Up to now, the analysis of aerial photos has been carried out by specialists who painstakingly examine every photograph, which makes it a very long, exacting and expensive job. The IMT-DMT of the EPFL and Antenne romande of FNP, aware of the special interest involved and the necessity of automated classification of aerial photos, have pooled their resources to develop a software program capable of differentiating between single trees, copses and dense forests. The developed algorithms detect the crowns of the trees and the surface of the orthogonal projection. Form the shadow of each tree they calculate its height. They also determine the position of the tree in the Swiss national coordinate thanks to the implementation of a numeric altitude model. For the future, we have the prospect of many new and better uses of aerial photos being available to us, particularly where isolated stands are concerned and also when evolutions based on a diachronic series of photos have to be assessed: from timberline monitoring in the research on global change to the exploitation of wooded pastures on small surface areas.

Airborne synthetic aperture radar (SAR)-derived soil moisture (m_v) estimates were evaluated using field measured and distributed hydrologic model simulated m_v values in the Reynolds Creek Experimental Watershed (RCEW), located in a semi-arid mountainous watershed in Southwest Idaho. A soil moisture inversion algorithm was developed by comparing L-HH SAR sigma⁰ (dB), corrected for surface roughness and topographic effects, with field measurements. The algorithm was found to be consistent with an algorithm developed in a previous study done on other rangelands. The corrected SAR-based soil moisture maps show relatively dry conditions at lower elevations and vice-versa, which is consistent with field observations. Change detection was found to give results similar to those from the corrected SAR, but it tends to give lower soil moisture values than the field measured or corrected SAR-based approaches. Out of the two SAR-based approaches used for the determination of m_v, change detection is simpler and adequate for the measurement of soil moisture changes. The corrected SAR approach is relatively difficult but it is better for the areas where absolute m_v measurements are required and SAR data under dry conditions are not available. The hydrologic model tended to overestimate surface m_v, which is what SAR senses. This was due to the failure of the model in accurately accounting for surface evaporation processes unless subsoil was relatively dry. In general, the overall trends in m_v, especially for 1989 and 1991, were reasonably represented but probably a little over the actual values. The soil moisture simulation model, corrected SAR, and ground measurements provided a unique mounds to study the moisture changes and distribution patterns in rangelands with an imaging SAR.

It is known that to increase the work rate of automated complexes for processing aerospace imagery intended for traditional topographic tasks there are used coherent optical-digital processors. Theoretical aspects of their work and traditional designs of functional and optical schemes are described in numerous publications in detail. Therefore, without analyzing well known designs, we shall discuss applications of original models of optical-digital processors for automated imagery interpretation.

Feature extraction is the most fundamental and important problem in satellite forest image recognition problem, but how to extract the feature is an important step towards solving recognition problems. The forest image features commonly consist of visual feature, statistical feature, transform field feature, and algebraic feature. The paper uses SPOT remote sensing forest images as samples, SVD (singular value decomposition) and GLCM (gray-level co- occurrence matrix) as the methods of feature extraction, it also compares and analyzes these two methods. The results of comparison show that GLCM is more effective in describing the visual feature and SVD in describing the internal feature. Forest image statistical feature can describe the image macroscopic features, such as the texture feature shape feature, etc. But SVD can describe the image internal features. The SVD spectral features synthesize the features of image at the pixel level. The combination of the two methods can be more effective in the forest image recognition and classification. We use both SVD and GLCM to extract remote sensing forest image features, and by the combination of these two methods we have got a better recognition of the ground surface forest of remote sensing images than before. Our word shows that the forest recognition rate reaches up to 95%.

Micropropagation is a sophisticated technique for the rapid multiplication of plants. It has a great commercial potential due to the speed of propagation, the high plant quality, and the ability to produce disease-free plants. However, micropropagation is usually done by hand which makes the process cost-intensive and tedious for the workers especially because it requires a sterile work-place. Therefore, we have developed a prototype automation system for the micropropagation of a grass species (miscanthus sinensis gigantheus). The objective of this paper is to describe the robotic system in an overview and to discuss the vision system more closely including the implemented morphological operations recognizing the cutting and gripping points of miscanthus plants. Fuzzy controllers are used to adapt the parameters of image operations on-line to each individual plant. Finally, we discuss our experiences with the developed prototype an give a preview of a possible real production line system.

Commercial vegetative propagation of floricultural crops requires the segregation of plant cuttings into categories based on size. The cuttings however must be graded when they are planted ('stuck'), at which time the grade of a cutting is not easy to determine. This paper reports on a system that learns to classify cuttings from being shown examples of images of cuttings that have been graded by a human expert. Based on the example set, the system learns to grade cuttings into categories. We report the results based on a set of 150 geranium plants that were graded by our system and compare the results to the performance of an expert grader.

A flexible and universal sorting system for potted plants has been developed and has recently been implemented in several greenhouses in the Netherlands. The system contains one or multiple color cameras, a lightning and transport unit, image processing hardware and specially developed software. The platform can be applied for sorting of several types of potted plants because of the implementation of learning techniques using statistical discriminant analysis and a neural network classifier. For most applications, because of different product characteristics or specific user requirements, an adjustment or optimization of the system is necessary. The newest successful implementation is the grading of begonia plants, using two color cameras from different angles, grabbing in total 7 images of one plant. Two extra features have been added to the system to be able to sort begonia: an improved plant variety recognition and automatic recognition of asymmetrical plants.

In recent work, we have demonstrated a prototype machine vision seedling inspection system which shows strong promise for automating production-line grading. Precise morphological measurements and accurate grade assignment require reliable identification of the seedling root collar location. The large variability of seedling morphology makes automatic root collar location the most challenging aspect of machine vision seedling inspection. This function is currently achieved using a heuristic algorithm which relies on many operator-controlled parameters to extract root collar location cues based on seedling shape. Artificial intelligence techniques, specifically, neural networks, have yielded excellent performance in similar pattern recognition applications. Neural networks were developed to locate the seedling root collar in digital images acquired by a machine vision inspection system. Several neural network architectures and input feature sets are evaluated. Input features consist of those used by the heuristic algorithm, plus additional features extracted from each line in the seedling image. The performance of several neural networks was superior to that of the heuristic algorithm. Good performance was achieved by networks which used local (single line) features along with normalized line number as inputs. A hierarchical network which took inputs from 15 lines over a 140-mm window provided improved performance in one case. The best networks identified the root collar location with an average error of less than 1 mm and an error standard deviation of 12 mm.

An automatic selective herbicide spraying system for weed control has been developed. When driving at 10 km/h, the system can spray weeds with a spatial resolution of about 10 cm². Weed plants as small as 10 mm² can be detected. Thus, weed plants are sprayed upon, whereas parts of the treated surface where no weed is present are left free of herbicide. In comparison to full-field treatments, a considerable amount of herbicide is saved in this way, with positive effects on treatment cost and environmental friendliness. The system consists of a series of parallel opto-electronic weed sensors scanning the surface, velocity sensors, an industrial PC for data processing, and herbicide sprayers mounted behind the weed sensors. Unlike in other selective spraying systems, the weed sensors do not employ the reflectance properties of weed, but the fluorescence properties. Although the sensor have been originally developed for weed detection, modification of the optics makes them suitable for the detection of a wide variety of objects and chemicals in agriculture and industry.

Spectral and morphological features were used to detect temperature induced stress on tomato plants. Top projected canopy area (TPCA) and profile were selected as morphological features and the reflectance of plant under side canopy (USC) and its average gray level were chosen as spectral features. Temperature regimes (day/night, 18/6 hours) 24/21 degrees Celsius, 21/18 degrees Celsius, and 19.5/16.5 degrees Celsius were used. Both spectral and morphological features were capable of detecting temperature stresses. Reflectance and gray level of plant USC correlated with average environment temperatures. The stress was detected after one week from occurrence based on both morphological and spectral features. However, stress was detected more clearly based on spectral features.

A significant reduction in the amount of pesticides applied in agricultural and biological systems could be achieved using spot spray technology. To accomplish this, advanced plant sensor systems must be developed that can accurately locate and identify weeds from crop plants in the field. Currently, both public and commercial efforts have concentrated on single element optical sensors based on key reflective elements of the plant and soil system. Machine vision or image analysis is being investigated as another possible tool in plant sensing. It may provide valuable optical design information for less expensive single-element sensors. Moreover, shape features and textural analysis already provides simple broadleaf-grass classification based on staged plant images. These have not been thoroughly field tested. Another approach is to test image analysis algorithms, using three-dimensional rendering of weed and plant canopy architecture under complex lighting regimes. What was essentially done was to extract plant shape and textural information, along with essential physiological data from actual photographic images and then reassemble them as a virtual plant in the computer. A dissection program was written in C and efficiently extracts and stores irregular leaf shape and texture data. A canopy architecture program was written in C and Media Cybernetics HALO^R graphics routines under DOS Expanded Memory on a personal computer. The plant simulation model consists of a three dimensional space where simulated light rays are generated as diffuse or speculative illumination. Plant surfaces are simulated with actual textural maps. The virtual plant is then manipulated to generate images that would be seen with machine vision. Computer simulated weed images were used to generate and test different fields of view sizes for evaluating how single element optical sensors would respond to composite leaf-soil reflectance.

This study describes machine vision procedures which are able to classify defective cherries from non-defective cherries. Defects can be divided into bruises, dry cracks, and wet cracks. Bandpass filters that enhance the intensity contrast between bruised and unbruised cherries are determined. An optimum combination of two wavelengths is identified at 750 nm (near-infrared range) and 500 nm (green range). An optimum single wavelength is identified at 750 nm. The image acquisition using these filters is described. Four detection methods using single view infrared images were studied. One method performed well in classifying cherries with bruises and wet cracks from non-defective cherries. One detection method using single view green images is studied. It performs well in classifying cherries with dry cracks from non-defective cherries. One detection method using infrared images and another using green images are used in combination to perform the detection on the entire surface of cherries. Two images, infrared and green, are taken from each of 6 orthogonal directions from the cherries. The integrated classifier misclassified 13% of non-defective cherries, 16% of bruised cherries, 0% of cherries with wet cracks, and 10% of cherries with dry cracks.

The objective of this project is to develop the hardware and software by which data on the state of a perishable commodity during transit are collected. At predetermined time intervals, the data are transmitted to a control center where the data are compiled and analyzed. The instantaneous state (shelf life) of the commodity is calculated from the data to provide support for the management of the shipment. The proposed system components include (a) sense and log physical factors within a confined store or container, (b) dispatch the logged data to a receiver, (c) analyze the data to interpret the current quality of the perishable, and (d) propose a course of action based on historical data. Canada exports alfalfa cubes to overseas and this material may become moldy during transport. Twenty containerized shipments of the cubes were instrumented to acquire data on temperature and humidities within each container. The data were correlated to the state of the material in terms of its moisture content and moldiness. The analysis of the thermal environment of the container showed that, for a satisfactory prediction of product temperature and moisture, the record of container wall temperature, headspace temperature, and the relative humidity of the container are required.

The possibility of using x-ray radiographic imaging for detecting intenal defects in apples has been investigated. Four hundred to seven hundred each of five Washington State cultivars [Red and Golden Delicious (RD< GD), Fuji (FJ), Granny Smith (GS) and Braeburn (BR)], both defect free and with assorted internal defects (bruises, senescence browning, rot, insect damage and watercore), were imaged using film and on-line line-scanning x-ray equipment. Both axial (stem-to-calyx) and radial images were obtained. The resulting images were presented to human operators, both as still shots and by scrolling them across the screen of a PC at rates approximating that of a commercial packing line. Good recognition [greater than 50% recognized, less than 10% false positives] could generally be obtained on selected cultivars when still shots were inspected. Apple orientation was required for recognition of watercore and rot. However, recognition rapidly fell off as scrolling speeds across the screens approached commercial rates. It is concluded that such inspection may be possible using machine recognition, but probably cannot be achieved using human operators.

In order to develop sorting methods for insect damaged pistachio nuts, a large data set of pistachio x-ray images (6,759 nuts) was created. Both film and linescan sensor images were acquired, nuts dissected and internal conditions coded using the U.S. Grade standards and definitions for pistachios. A subset of 1199 good and 686 insect damaged nuts was used to calculate and test discriminant functions. Statistical parameters of image histograms were evaluated for inclusion by forward stepwise discrimination. Using three variables in the discriminant function, 89% of test set nuts were correctly identified. Comparable data for 6 human subjects ranged from 67 to 92%. If the loss of good nuts is held to 1% by requiring a high probability to discard a nut as insect damaged, approximately half of the insect damage present in clean pistachio nuts may be detected and removed by x-ray inspection.

X-ray images of pistachio nuts on conveyor trays for product inspection are considered. The first step in such a processor is to locate each individual item and place it in a separate file for input to a classifier to determine the quality of each nut. This paper considers new techniques to: detect each item (each nut can be in any orientation, we employ new rotation-invariant filters to locate each item independent of its orientation), produce separate image files for each item [a new blob coloring algorithm provides this for isolated (non-touching) input items], segmentation to provide separate image files for touching or overlapping input items (we use a morphological watershed transform to achieve this), and morphological processing to remove the shell and produce an image of only the nutmeat. Each of these operations and algorithms are detailed and quantitative data for each are presented for the x-ray image nut inspection problem noted. These techniques are of general use in many different product inspection problems in agriculture and other areas.

Classification results for agricultural products are presented using a new neural network. This neural network inherently produces higher-order decision surfaces. It achieves this with fewer hidden layer neurons than other classifiers require. This gives better generalization. It uses new techniques to select the number of hidden layer neurons and adaptive algorithms that avoid other such ad hoc parameter selection problems; it allows selection of the best classifier parameters without the need to analyze the test set results. The agriculture case study considered is the inspection and classification of pistachio nuts using x- ray imagery. Present inspection techniques cannot provide good rejection of worm damaged nuts without rejecting too many good nuts. X-ray imagery has the potential to provide 100% inspection of such agricultural products in real time. Only preliminary results are presented, but these indicate the potential to reduce major defects to 2% of the crop with 1% of good nuts rejected. Future image processing techniques that should provide better features to improve performance and allow inspection of a larger variety of nuts are noted. These techniques and variations of them have uses in a number of other agricultural product inspection problems.

Decreasing cost of computer technology has made it feasible to incorporate machine vision technology into the agriculture industry. The biggest attraction to using a machine vision system is the computer's ability to be completely consistent and objective. One use is in the variety discrimination and quality inspection of grains. Algorithms have been developed using Fourier descriptors and neural networks for use in variety discrimination of barley seeds. RGB and morphology features have been used in the quality analysis of lentils, and probability distribution functions and L,a,b color values for borage dockage testing. These methods have been shown to be very accurate and have a high potential for agriculture. This paper presents the techniques used and results obtained from projects including: a lentil quality discriminator, a barley variety classifier, a borage dockage tester, a popcorn quality analyzer, and a pistachio nut grading system.

Intrinsic end-use quality of hard winter wheat breeding lines is routinely evaluated at the USDA, ARS, USGMRL, Hard Winter Wheat Quality Laboratory. Experimental baking test of pup loaves is the ultimate test for evaluating hard wheat quality. Computer vision was applied to developing an objective methodology for bread quality evaluation for the 1994 and 1995 crop wheat breeding line samples. Computer extracted features for bread crumb grain were studied, using subimages (32 by 32 pixel) and features computed for the slices with different threshold settings. A subsampling grid was located with respect to the axis of symmetry of a slice to provide identical topological subimage information. Different ranking techniques were applied to the databases. Statistical analysis was run on the database with digital image and breadmaking features. Several ranking algorithms and data visualization techniques were employed to create a sensitive scale for porosity patterns of bread crumb. There were significant linear correlations between machine vision extracted features and breadmaking parameters. Crumb grain scores by human experts were correlated more highly with some image features than with breadmaking parameters.

The food industry has begun using lasers for improving the effectiveness of production. We clearly see three main directions: in food technology, in food engineering, and in agriculture. The resulting work of the seminar with concrete results, gotten by some research centers, is given. The programs of the seminar for 1988-1996 also are given.

This report presents the results of theoretical and experimental investigations of a new type of electro- discharged carbon-dioxide laser. It has been shown that electrical power coupling in discharge can reach the level of 300 W/cm³. It is more than two order higher than in ordinary gas discharge. The relations of similarity and scaling lows for vortex electro-discharged carbon-dioxide laser has been proposed. This laser can be used for various biological, agricultural and forestial processing.

This work concerns luminescence of egg shell and chick down and livability and egg production of hens. It was established that the best layers showed orange color of egg shell luminescence at the beginning of the laying period. Yellow color of chick down luminescence indicates on the chicken's good development especially concerning its digestive and circulatory systems, and connects with its future high livability and egg production. So the use of the color of luminescence is advisable as an additional characteristic providing good possibility to forecast the development of a chicken, its resistance for stressors, its livability and egg production afterwards.

Artificial vision and image analysis are increasing their role in agriculture. Using systems based on stereoscopic vision it is possible to associate to the large images information, a three dimensional space reference. So it is possible to measure distances between vision system and any point of real observed scene or calculate relative positions between different subjects of the same image. The work evaluates the possibility, the capacity and the accuracy of stereovision system as possible application in environmental and agricultural survey. The analysis was performed theoretically in function of the characteristics of some image acquisition CCD equipment, existing on the market of video-photographic device, and considering different parameters of environmental situations (field of view width, linear distance - z - between video system and the subject). Good accuracy is obtainable also by a 'standard' system (500 pixels resolution) for a z distance of 100 m and 5 m distance of the two video equipment. For a similar situation with a high performance equipment (3060 pixel resolution), it is possible to obtain an accuracy of the centimeter rank.

Visual quality control in many food processing operations continues to be a manual, difficult, and tedious task. Computer/machine vision systems offer a solution, but the development of effective algorithms that are able to accommodate the natural variability of food products has proved to be problematic. This paper examines and compares three techniques for processing multi-spectral imagery for these applications. One technique is to use artificial neural networks (ANNs). ANNs have the ability to be fault tolerant when establishing decision surfaces within the test data and can operate in parallel at high speeds -- this makes them ideal for this application. The main drawback of ANNs is their inability to provide a meaningful justification for the decision boundaries they establish when classifying data. Another image processing technique that uses a more deterministic data classification method is vector quantization (VQ). VQ uses a data clustering and splitting algorithm that can be modified to improve speed and accuracy according to the application. In an effort to include all levels of algorithm complexity, a modified thresholding approach is also compared to the more computationally demanding ANN and VQ techniques. The strengths and weaknesses of each of these algorithms are highlighted based on their performance in these domains.