Image filtering and restoration using Two Dimensional fast Hartley transformation is analyzed on a PC-386 personal computer. Images are captured using a PC-based frame grabber for subsequent filtering. To speed up data transfer during processing, data is stored in expanded memory of the computer. Application of fringe pattern enhancement in moire methods for vibration studies are demonstrated.

The reflection moire is a well known method to analyze the out-of-plane slope and curvature of a plate. It has also been applied to obtain the mode shape of a plate vibrating at resonance. The sensitivity of the method depends on the pitch of grating and the set-up parameters. In order to increase the sensitivity, the set-up or pitch of grating needs to be changed. Besides that, patterns produced with ordinary gratings have to be filtered to increase contrast. Using computer grating with image processing, the gratings can be quickly and readily changed and images can be enhanced. In this paper, the reflection moire method using computer grating is proposed to inspect the mode shapes of vibrating plates. A method to calculate the frequency of a vibrating plate using scanning frequency of screen is also introduced. Some of the routines for fringe enhancement are quasi-real-time while other routines are more computationally intensive.

A focus of attention mechanism for locating blob- and edge-like visual steel surfe defects, is proposed in the paper. An inherent problem with focus of attention mechanisms is that weak events arc masked by strong events and pass unnoticed. Solutions to this problem, based on model driven purposive vision, arc discussed. Results obtained with the attention mechanism are illustrated with sample fect images of stretcher streams, roll marks and rust spots.

An algorithm of motion estimation using straight line correspondences and its error analysis are presented in this paper. First, a linear algorithm of motion estimation is briefly reviewed. The input data of the algorithm are point set correspondences from which line correspondences are derived. At least thirteen line correspondences are needed for the algorithm. Then, a method for error analysis for this algorithm is proposed. The error analysis is statistical and based on first order perturbation. Some simulations are carried out. The relationship between the errors in the motion estimation results and the 3D motions are also investigated.

This paper describes two methods for the capture of the 3-D co—ordinates of an object for use in automated CAD—CAM system. The workings and details of the familiar structured light method and the wholefield projection grid method are described.

A focus of attention mechanism for locating blob- and edge-like visual steel surfe defects, is proposed in the paper. An inherent problem with focus of attention mechanisms is that weak events arc masked by strong events and pass unnoticed. Solutions to this problem, based on model driven purposive vision, arc discussed. Results obtained with the attention mechanism are illustrated with sample fect images of stretcher streams, roll marks and rust spots.

Over the last decade or so, several parallel advances have been made in the two distinct disciplines: the fuzzy logicand the neural networks. As the names imply, the fuzzy logic provides a mathematical strength to the emulation of certain perceptual and linguistics attributes associated with the human cognition, whereas the science of neural networks provides new computing morphologies with learning and adaptive capabilities. A marriage between these two distinct disciplines have the potential of producing off-springs with the capabilities of generating a new discipline -the virtual intelligence (VI) with a new generation of computing systems - the virtual cognitive systems (VCS). Such a virtual cognitive systems will, hopefully, recapture certain aspects of human cognition. An integration of these two fields: the fields of neural networks and fuzzy logic has a potential of producing robust sensors and robust control mechanisms. In this paper, we briefly explore the possibilities of such a new system -the virtual cognitive systems, in the service of humanity.

An experiment in determining 3D vehicle motion in an outdoor scene using line correspondences over an image sequence is presented in this paper. The motivation of this experiment is to test how well algorithms of motion estimaLion work on real life outdoor scenes. Using images taken by a well calibrated camera, the experiment includes straight line extraction from intensity images, straight line matching and motion estimation using line correspondences. Many corrections and adjustments are performed on the raw image data before they are used for motion estimation, which include film distortion correction, lens distortion correction and camera alignment. The results of this experiment indicate that for the type of vehicle motion under consideration, our algorithm can obtain reasonably accurate estimates for rotation. The experiment also reveals that there is a big gap between simulation tests and real scene tests for algorithms of motion estimation.

We present a method and its optical and digital processor for the detection of defects in glazed china bone plates. The method consists in exploring the properties of non-coherent plane polarized light to enhance optical defects whose diameter is larger than the maximum human eye resolution (0. 1 mm). Changes in glazing smaller than 0. 1 mm are not considered to be defects. The image is digitized after collection by a CCD camera. The board sees 8 bits of gray level and it has 512x512 pixels. The digitized image is further processed. We have employed morphological analysis in order to establish a figure of merit for each plate. The merit function is related with the number of defects, their color, and size. The quality classification in then readily obtained. The algorithms were implemented in a hardware configuration which is briefly outlined.

This paper describes an automatic visual inspection algorithms for complex printed pattern. The inspection algorithms are based on a framework of structural approach for feature verification. It consists of a pre-processing algorithm based on low level morphological thinning operators and a model-based analysis algorithm. In addition, a width checking algorithm to detect width violations is also proposed. All the methods presented were tested on flexible printed circuit boards (PCBs) as an example.

This paper describes an algorithm for detecting walls and openings in a maze from an image captured through a camera. The captured image is first transformed from 256 gray levels to a binary one, a procedure which extracts the edges of the walls of the maze. Using the bilinearly distorted image as a guide, a coordinate system is developed so that the pixel value of the critical points within the maze could be retrieved to determine whether the points represented a wall or an opening. The results are output in an array which gives the configuration of the maze as seen by the image algorithm developed. Given a well-thresholded image, the processing time is 38 seconds and the level of accuracy obtained is 86%.

This paper presents a real time inspection algorithm of fuel pellet surface. The algorithm identifies the different types of surface features of the pellets from sensed data derived from an array of proximity sensors. The algorithm meets the operating requirements in terms of speed and accuracy and can be implemented using relatively simple hardware. The development also includes a sensed data visualization tool which facilitates the analysis of the performance of the algorithm.

In this paper, we propose a syntactic pattern recognition method for non-schematic drawings, based on a new attributed graph grammar with flexible embedding. In our graph grammar, the embedding rule permits the nodes of a guest graph to be arbitrarily connected with the nodes of a host graph. The ambiguity caused by this flexible embedding is controlled with the evaluation of synthesized attributes and the check of context sensitivity. To integrate parsing with the synthesized attribute evaluation and the context sensitivity check, we also develop a bottom up parsing algorithm.

Monitoring of the cutting process is one of the most crucial and difficult problems in unmanned machining technology. In this paper, the effects to the thrust and the torque in drilling by some abnormal cutting states have been investigated. Experiments show that the dynamic components of the thrust and the torque provide useful information for identifying different tool failures. Reasonable explanations have been given for the phenomena from physical mechanisms. A mathematical model for the dynamic cutting process was proposed and some problems about the estimation of the time series AR(n) model were also discussed. After having compared several defined monitoring parameters, a new inprocess signal processing method, `adaptive variance-residual recursive algorithm,' based on least squares theory has been suggested with which signal signatures of cutting states can be extracted effectively. It bears the advantages of rapid response, less amount of calculation, and easy to set deterministic thresholds. The in-process monitoring system developed shows a bright prospect in practical application.

A hybrid pyramid multiprocessor vision machine has been built for automated object recognition which is vital for manufacturing automation. The machine consists of 512 one- dimensional single-instruction multiple-data (SIMD) array processors at the bottom and a 63- node transputer-based multiprocessor system on the top. This paper describes a Hierarchical Linear Generalized Hough Transform (H-LIGHT) and its implementation on the SFU hybrid pyramid. It is shown that the hierarchical algorithm improves the robustness of the pattern matching process and increases the efficiency in detecting objects in all possible orientations. It has great potential for real-time pyramidal implementation.

This paper presents different methods for the automatic measurement of industrial parts by means of an artificial vision system. Performances in relation with precision and computer time consumption are discussed. Different algorithms are presented on industrial study cases in order to perform accurate measurements. Most methods work on grey level image and use subpixel algorithms. In this case, the precision is less than the traditional ratio `image size divided by the number of pixels.' In practice the problem is not so simple and even performing algorithms are not strong enough when lighting conditions are not optimum. For this reason, it's interesting to compare different study cases. Different hardware solutions from a turn key system to an image board for a personal computer have been used and we compare the set-up time which has been required in order to realize the complete vision system in both solutions.

Three dimensional (3-D) digital images and patterns under transformations are facilitated by the splitting-shooting method (SSM) and the splitting-integrating method (SIM). The combination (CSIM) of SSM and SIM and the combination (CIIM) of SIM and SIM are proposed for a cycle conversion T^-1T, where T is a nonlinear transformation, and T^-1 is its inverse transformation. This paper focuses on exploitation of accuracy of the obtained image greyness. In our discrete algorithms, let a 3-D pixel be split into N³ subpixels. The convergence rates of sequential errors can be denoted by O(1/N^k), k >= 1. High convergence rates indicate less CPU time needed. Both error bounds and computation of pixel greyness have shown the following important conclusions: (1) O(1/N) for CSIM; (2) O(1/N) or O(1/N²) for CIIM; (3) O(1/N³) for CIIM using quadratic B-spline functions in antialiasing images.

This paper presents the techniques used to inspect for defects on surface mount PCBs. Focusses were made on five different types of defects, namely, missing components, misalignment, wrong orientation of I.C. chips, wrong parts, and poor solder joints. Thus, five separate algorithms had been developed to detect these faults. The technique of windowing was employed to reduce the amount of redundant data to be processed. Preprocessing functions like convolution as well as all image processing tasks are performed on the window regions only, saving tremendously on computation time. Experimental results showed that these algorithms are reliable, fast, and cost-effective.

Two commonly used optical correlation techniques, matched spatial filtering and joint-Fourier transform correlation, are briefly reviewed. A recently proposed real-time joint-Fourier transform correlation is then discussed and demonstrated by computer simulations.

A robust estimator, called simplified MF-estimator, to determine 3-D motion parameters is investigated in the paper. According to the Bayesian decision rule, it partitions a given data set into a subset containing good observations and a subset containing bad observations. The estimator can implement a least-squares estimator on the good data, and down-weight the outliers. To speed convergence of the algorithm, an annealing schedule is used. Finally, a great number of simulations are conducted to show robustness of this algorithm.

The difficulties commonly encountered in developing laser imaging technologies are: (1) high cost of the laser system, and (2) time and cost involved in modeling and maneuvering a physical environment for the desired scenes. In contrast to the real imaging systems, computer generated laser images provide researchers with fast, accurate, cost-effective data for testing and developing algorithms. The laser imaging simulation package (LISP) described in this paper provides an interactive solid modeler that allows users to construct the artificial environment by various solid modelling techniques. Two fast ray tracing algorithms were developed and discussed in this paper for generating the near realistic laser data of any desired scene. These computer generated laser data facilitates the researchers in developing laser imaging algorithms. Thus, LISP not only provides an ideal testbed for developing and testing algorithms, but also an opportunity to explore the limitation of laser imaging applications.

The paper presents the recent intelligent robotics research being carried out at the PAMI Lab of the University of Waterloo, Waterloo, Ontario. The intelligence control of manipulators is directed and guided by 3-D vision. It is implemented for a mobile robot and robot manipulators in a workcell. The intelligent robotic system is capable of: (1) real-time recognition and location of 3-D objects and obstacles with a single camera system mounted on the robot arm; (2) optimal trajectory planning for a robotic manipulator with obstacle and singularity avoidance capability; and (3) vision directed navigation of a mobile robot. Application of this technology to industrial and space station projects is included in the discussion.

A neural network system to automatically inspect the defects of textured objects is presented. The system is composed of two parts, the first part consists of a self-organizing neural network that's main task is to segment the image into different regions which are of different texture characteristics; the second part consists of a neural network that's architecture is similar to that of a Boltzmann machine. Its main goal is to restore the image with defects to a perfect one. All defects can be detected by simply comparing the two images. Our method is optimum for textured objects, because it needs neither the precise registration which is necessary for many inspection methods and difficult to realize, nor the image pre-treatment which is always time consuming.

In the automatic rendezvous and docking manoeuvre (RVD) of space activity, determining the 3-D location and attitude between two vehicles is most important. A vision system to perform the docking manipulation in RVD is described in this paper. An improved algorithm is used for calibrating the geometric parameters of a camera fixed on the tracking vehicle off-line. Because the line-off-sight angles of four markers on the target vehicle to the lens center of the camera can be computed according to the optical principle and vector theory, the locations of the vehicle are obtained from the solution for a set of nonlinear equations from the triangular theory. The attitude angles for the vehicles are solved by a translational matrix of target frame to vehicle frame. As the vehicle closes in to the target, sets of markers having different distance intervals or a list of calibration parameters for cameras with different fields of view are selected at the proper moment to improve the situation when at least one of the markers exceeds the field of camera view. The series of experiments is given. The vision system is run on a SUN-4/330 Sparc station system equipped with one image board IT-151 and a CCD TV camera. All software is written in C language.

This paper presents the development and realization of a robotic vision system for the recognition of 3-dimensional (3-D) objects. The system can recognize a single object from among a group of known regular convex polyhedron objects that is constrained to lie on a calibrated flat platform. The approach adopted comprises a series of image processing operations on a single 2-dimensional (2-D) intensity image to derive an image line drawing. Subsequently, a feature matching technique is employed to determine 2-D spatial correspondences of the image line drawing with the model in the database. Besides its identification ability, the system can also provide important position and orientation information of the recognized object. The system was implemented on an IBM-PC AT machine executing at 8 MHz without the 80287 Maths Co-processor. In our overall performance evaluation based on a 600 recognition cycles test, the system demonstrated an accuracy of above 80% with recognition time well within 10 seconds. The recognition time is, however, indirectly dependent on the number of models in the database. The reliability of the system is also affected by illumination conditions which must be clinically controlled as in any industrial robotic vision system.

In many potential machine vision applications, the biggest constraint to its industrial use is the cost of the image processing hardware. In fact, there are many visual tasks performed by human operators that could be easily automated. However, it is very difficult to find in the market a vision system with the adequate relation between cost and performance. Most of the image processing systems need too much time to process one image, or have a very high cost. We present an architecture that was developed for surveillance purposes and that we are applying to several applications, namely to the semi-conductor industry and to flexible manufacture. The hardware was conceived in a modular approach with real time performance in each module. Cascading of several modules can produce different image processing functions with an image or data, being generated in every frame of video. The low cost of the hardware together with its very high performance, allow many different industrial applications in particular in the field of industrial automation. The results obtained with the industrial prototypes are presented in two applications and other possible applications are in progress.

The paper industry uses wood as its raw material. To know the quantity of wood in the pile of sawn tree trunks, every truck load entering the plant is measured to determine its volume. The objective of this procedure is to know the solid volume of wood stocked in the plant. Weighing the tree trunks has its own problems, due to their high capacity for absorbing water. Image processing techniques were used to evaluate the volume of a truck load of logs of wood. The system is based on a PC equipped with an image processing board using data flow processors. Three cameras allow image acquisition of the sides and rear of the truck. The lateral images contain information about the sectional area of the logs, and the rear image contains information about the length of the logs. The machine vision system and the implemented algorithms are described. The results being obtained with the industrial prototype that is now installed in a paper mill are also presented.

The production of cork stoppers is based on a manual punching process. A skilled operator performs a rapid visual inspection of the raw material to determine the location of defects and then punches the stoppers one at a time, avoiding those defects which he considers detrimental to the stoppers. This paper presents a machine which has been developed to punch stoppers in a completely automatic way. The machine employs a vision and image processing system which can detect defects and generate a corresponding punching pattern in a way which emulates the conventional manual process. Two monochrome CCD cameras and an image processing board are used to acquire and process the necessary visual information about the raw cork material. The algorithms implemented in the vision system are described, together with the image processing hardware. First results obtained with the machine are presented and a comparison with the manual method is made. This development work was carried out under a contract between LNETI, a government research center, and two private companies, EID and Mecanova.