The inspection of fine pitch surface-mounted devices by comparison of defect-free and defective packages is a promising area of research. The types of defects considered include missing pins, bent pins, broken pins, and bad solder connections on mounted packages. The detection algorithm includes morphological image processing operations followed by a neural network. The feature extraction steps include morphological filtering for thresholding, skeletonization, and centroid determination. The centroids are used as inputs to a backpropagation neural network for determining the presence of defects. The neural network compares the input data against data representing defect-free packages and produces a measure of how closely the two data sets match. The accuracy of the network in identifying both good and defective packages is discussed with output values interpreted both incorporating and not incorporating rejection on the bases of a minimal threshold for output values and a minimal separation between output values. The algorithm performance is evaluated based on its performance in correctly identifying the presence or absence of a number of frequently found defect types. Evaluation is also based on the neural network performance with different training parameters. The neural network is shown to identify defects over 70% of the time without rejection and often 100% of the time with rejection.
Skeletonization of binary images is an essential step in the inspection of many products, most notably printed circuit boards. It also is used in many other situations, an unusual example being the location of branching points on growing plants for purposes of cutting and vegetative propagation. Commercially-available image processing boards typically can't perform this operation, although they readily perform the easier task of repeated binary erosion. While a single skeletonization step cannot be done in one pass using a 3 X 3 neighborhood, one pass with a 4 X 4 neighborhood suffices. This result has been implemented in custom integrated circuits imbedded in proprietary products, but (to our knowledge) is not commercially available. This paper describes a new pipelined implementation of binary skeletonization which fits easily into the standard SKIPSM (Separated-Kernel Image Processing using finite State Machines) architecture and which can be built using standard ICs costing less than $DOL200 total. The same approach also can be implemented in software, providing an order-of-magnitude increase in speed at no extra cost. Furthermore, this same SKIPSM architecture is highly versatile and programmable, allowing it to be software- reconfigured to perform hundreds of other pipelined image processing operations.
The automatic, real-time visual acquisition and inspection of VLSI boards requires the use of machine vision and artificial intelligence methodologies in a new `frame' for the achievement of better results regarding efficiency, products quality and automated service. In this paper the visual detection and classification of different types of defects on solder joints in PC boards is presented by combining several image processing methods, such as smoothing, segmentation, edge detection, contour extraction and shape analysis. The results of this paper are based on simulated solder defects and a real one.
A generic approach to the development and integration of machine vision, within surface- mount electronics manufacturing, has been proceeding based on the concept of a standard vision framework. A framework is a collection of system components, the connection of which can be configured with appropriate support tools. This is facilitated using object- oriented analysis and design techniques to identify and describe those elements, or modules, that are crucial to all vision systems within the domain. Analysis of surface-mount manufacturing has identified fifteen potential tasks in which machine vision inspection and control is beneficial. The essential functionality which spans these tasks has been identified and incorporated in a set of approximately twenty visual components implemented using the KAPPA programming environment. A practical exploration has been made into using the framework to develop a method of classifying insufficient solder deposits based on the distinct light reflection characteristics of solder fillets when illuminated from different angles. Classification has been reliably achieved by calculating the variation in mean luminance of specific fillet regions between images obtained with high and low angles of lighting using a custom light source. The resulting system architecture has illustrated the potential of object- oriented software and specification techniques, producing an elegant structure based on code reuse and `design by extension'.
This paper introduces algorithm design concepts that have proven useful in real-time X-ray solder joint inspection systems. Real-time industrial inspection systems face tremendous challenges, which are not always obvious when first designing such systems. The goal of this paper is to introduce a few of these challenges, as well as software and algorithm methodologies that effectively deal with some of these challenges. Specific examples included involve the real-time inspection of circuit board solder joints using cross-sectional X-ray images. Methods for dealing with solder joint orientations, utilities preserving subpixel and subgray level accuracy, and designs enabling code sharing are presented.
This paper summarizes the research activity on real time landmark tracking being developed at Robotics Laboratory at Pavia University. A visual tracking system for mobile robot PARIDE navigation is described. The purpose of the tracking system is to keep in the field of view "relevant" objects lying in the robot surroundings. The objects to be tracked are those which have some significance that can be used for extracting information about the surrounding environment or about the robot position. In particular the paper focuses on image processing and landmark design. Motion control analysis, algorithms and a general framework description can be found in  and [1 1]. Software development has been constrained by real time requirements imposed by the mobile robot kinematics and on environment hypothesis. The algorithm has been tested on PC based platform in order to speed up the development setting and to test different alternatives in short time.
Keywords: real time vision, active vision, landmark tracking, mobile robots
An optoelectronic target tracking system based on a BPEJTC is presented. It is shown that since the BPEJTC provides higher peak-to-sidelobe ratio than any other versions of JTC and does not cause correlation peaks due to intra-class association, this system is well-adaptive to a multi-target tracking environment. Moreover, this system effectively distributes required computations in that repeated time-consuming Fourier transforms are taken in the optical unit of the system while some arithmetic manipulations are conducted in the digital unit. In addition to the designed system architecture, some experimental results conducted by this system are illustrated.
Industrial automation, robotics and automatic inspections include the processing of image sequences. Especially the registration of dynamic process parameters such as velocity and acceleration is realized with real time image processing. In this paper a visual motion detection system is used for contactless measurement of dynamic process parameters in a continuous rolling process. The motion of hot tubes is detected during an image sequence analysis by observing the translation of specific templates that are extracted from the texture of the tubes' surfaces. The texture is necessary for faultless operation of the system. Actually there are some tubes without a clearly visible texture. The paper presents methods for preprocessing the image's data in order to model pseudo texture. The motion is analyzed by using correlation procedures in several areas of interest in an image. The paper describes the problems of an on- line motion detection in an image sequence and offers theoretical and practical approaches to solve them.
In this paper, we make an attempt to predict the location of a cable in next frame according to some parameters in current frame. At first, the initial condition is that the approximate location of cable must be given. From Hough Transformation, a high accumulated degree point in Hough Space is gotten. Afterwards, by inverse Hough Transformation, using that point, the location is detected and then cable is pointed out from its original image. Based on actual working conditions and performance indices, as well as the location in current frame, the maximum range of degree within which the cable may occur in next frame with high possibility is evaluated. Furthermore, a narrow range of location in next frame is confirmed according to the speed of robot. It is this narrow range that reduces the influence of the background on the detected object to the least. Therefore, the location of cable can be detected more accurately. Still more, as the final target of this project, an approach to detect the dangerous status of Cable is shown. In this approach, we want to distinguish the predict area into three parts with a simple area partition. By means of time sequence, the relative changes on the size of area which is between shadow and cable can conclude whether the cable is suspended in water.
This paper presents a statistical method for classification of bottles in crates for use in automatic return bottle machines. For the automatons to reimburse the correct deposit, a reliable recognition is important. The images are acquired by a laser range scanner coregistering the distance to the object and the strength of the reflected signal. The objective is to identify the crate and the bottles from a library with a number of legal types. The bottles with significantly different size are separated using quite simple methods, while a more sophisticated recognizer is required to distinguish the more similar bottle types. Good results have been obtained when testing the method developed on bottle types which are difficult to distinguish using simple methods.
While the uniform sampling method is quite popular for pointwise measurement of manufactured parts, this paper proposes three novel sampling strategies which emphasize 3D non-uniform inspection capability. They are: (a) the adaptive sampling, (b) the local adjustment sampling, and (c) the finite element centroid sampling techniques. The adaptive sampling strategy is based on a recursive surface subdivision process. Two different approaches are described for this adaptive sampling strategy. One uses triangle patches while the other uses rectangle patches. Several real world objects were tested using these two algorithms. Preliminary results show that sample points are distributed more closely around edges, corners, and vertices as desired for many classes of objects. Adaptive sampling using triangle patches is shown to generally perform better than both uniform and adaptive sampling using rectangle patches. The local adjustment sampling strategy uses a set of predefined starting points and then finds the local optimum position of each nodal point. This method approximates the object by moving the points toward object edges and corners. In a hybrid approach, uniform points sets and non-uniform points sets, first preprocessed by the adaptive sampling algorithm on a real world object were then tested using the local adjustment sampling method. The results show that the initial point sets when preprocessed by adaptive sampling using triangle patches, are moved the least amount of distance by the subsequently applied local adjustment method, again showing the superiority of this method. The finite element sampling technique samples the centroids of the surface triangle meshes produced from the finite element method. The performance of this algorithm was compared to that of the adaptive sampling using triangular patches. The adaptive sampling with triangular patches was once again shown to be better on different classes of objects.
This paper describes a computationally efficient matching method for inspecting 3D objects using their serial cross sections. Object regions of interest in cross-sectional binary images of successive slices are aligned with those of the models. Cross-sectional differences between the object and the models are measured in the direction of the gradient of the cross section boundary. This is repeated in all the cross-sectional images. The model with minimum average cross-sectional difference is selected as the best match to the given object (i.e., no defect). The method is tested using various computer generated surfaces and matching results are presented. It is also demonstrated using Symult S-2010 16-node system that the method is suitable for parallel implementation in massage passing processors with the maximum attainable speedup (close to 16 for S-2010).
Quality control in switch manufacturing requires both an electrical test and full analysis of switch's plastic surface to mount. This paper describes an automatic system for stain and dent fault-detection for the plastic surface of switches during its manufacturing. Two different images of every switch are used: in the former, the switch is illuminated with fuzzy and uniform white light intended to magnify the visual effects of the stain; in the latter, a known striped pattern is projected on the switch so that a dent or a bump at the plastic surface will result in a modified pattern that would easily be detected by the system. The inspection unit is designed to operate as a real-time and low-cost system. So, a single inspection unit takes and processes both pictures. Furthermore, another operation is performed for unfaulty parts that pass the inspection test: they are properly oriented for further assembling processes based on their serigraphied symbols. Some improvements could be made to the system at the expense of low-cost constraints that involve better accuracy and provide with more robustness in few specific cases.
This paper presents a system that integrates a CAD package and a vision system to attain a higher level of intelligence and automation in manufacturing environments. The proposed vision system can achieve object recognition without tedious, manual teaching in advance by adopting a simple recognition method and the use of CAD data. It can also on-line derive the 3D position and orientation of the workparts randomly placed on the workbench from their camera images, for guiding robot to pick them up. Such derived object position and orientation data can also be used to compensate the workpart measurement path coordinates of coordinate measuring machine (CMM). The workpart searching ability over the workbench has been demonstrated in this proposed system as well. The CAD package functions are extended for not only being used in product design, but also in producing simulated camera images to correspond to those from the real camera in the robot workcell simulation. With these developments, a higher level intelligent and automatic manufacturing environment without the need of accurate part feeder, fixture, and auxiliary toolings is achieved.
Key Words: Vision system, CAD images, Object recognition, CMM
A range image is an image where each pixel represents a measurement of the distance from the camera to the object. Typical applications for range imaging are inspection and dimension measurement in industrial processes, e.g. in the forest products industry. Range image acquisition can be made in many different ways. In this paper we use an active triangulation method where a sheet-of-light illuminates the scene. The sensor level signal processing task is to extract the light impact position in each sensor column. Two novel algorithms implemented on the commercially available smart image sensor MAPP2200 are presented. Both algorithms give 256 pixel width resolution at a line frequency of 15 - 20 kHz, corresponding to range pixel rates of 4 - 5 MHz, and range resolution varying from 8 up to 13 bits in special cases. This is considerably faster than other proposed methods. One of the methods also gives intensity data concurrent and in perfect registration with the range data and the other has an error detection capability to detect multiple peaks on the sensor.
A system is described that will give the range to different portions of a scene. The optical system is modified such that the optical transfer function has periodic nulls, with the period of the nulls depending on the axial distance within the volume of the 3D scene. These same nulls appear in the spatial frequency spectrum of the images of objects that are at different distances within the volume of the scene. Estimation of the period of the nulls in the spatial frequency spectrum of the image then gives the ranges to different objects within the scene. A similar optical/digital processing technique is used to extend the depth of field of the composite system so that it will work over larger changes in range.
Time Delay and Integration imaging offers a complete solution to the peripheral inspection/imaging of rotating cylindrical objects. Coupled with simple structured light schemes, the deformation or surface contour of the cylindrical object is highlighted and quantified. High speed TDI facilitate inspection of fast rotating objects, a feature preferred in industrial inspection systems. Experiments presented here are performed at rotation speeds of upto 2500 RPM. The experimental setup, influence of various system parameters are discussed in this paper. Examples using a food powder can as the object is provided.
Keywords: Machine vision, Digital imaging, Visual inspection, Moire methods
This paper describes a typical machine vision system in an unusual application, the automated visual inspection of a Casino's playing tables. The SORTE computer vision system was developed at INETI under a contract with the Portuguese Gaming Inspection Authorities IGJ. It aims to automate the tasks of detection and classification of the dice's scores on the playing tables of the game `Banca Francesa' (which means French Banking) in Casinos. The system is based on the on-line analysis of the images captured by a monochrome CCD camera placed over the playing tables, in order to extract relevant information concerning the score indicated by the dice. Image processing algorithms for real time automatic throwing detection and dice classification were developed and implemented.
The accuracy of the visual inspection process that performs the quality classification of lumber is one of the key interest areas in the mechanical wood industry. In principle, the quality classification of wood is straightforward: the class of each board depends on its defects and their distribution, as defined by the quality standard. However, even the appearance of sound wood varies greatly and there are no two boards or defects that have exactly the same properties such as color and texture. We describe the development of a color vision technology for grading softwood lumber. Much attention has been given to the early and cheap recognition of sound wood regions, as only a minor portion of the surface area of boards, around 5 - 10%, is defective. The non-interesting regions can be discarded and the hardware and communication bandwidth requirements at later defect identification stages are relieved. In the end the description of the board and its defects is passed to a grader that searches for all the applicable quality classes from the given set of standards. Extensive comparative tests have been carried out in a complete simulated system. The effects of changes in the spectrum of illumination have been evaluated to identify robust color features and to produce the requirements for color calibration.
The use of wastepaper as a raw materials within the paper making industry is steadily increasing. One of the key elements in processing wastepaper is the removal of ink. Current research in the industry is focused on the capability of measuring particle size on-line, within a paper mill environment. This paper describes the development of a prototype machine vision system for the analysis of ink particles within wastepaper pulp samples. A priori knowledge of the domain has been exploited to produce a lighting and sample presentation system which maximizes ink particle contrast within the image. The system has been used to examine samples from a pilot deinking cell at various stages of operation. This gives a wide spectrum of sample compositions typically encountered within a mill environment. One of the key areas in a fully automated analysis system is the segmentation of the images to determine ink particle content. Simple thresholding techniques provide the fast solution required to analyze the high throughput of a deinking line. Due to the large variation in image content the development of an overall segmentation strategy incorporating several algorithms is proposed. A suitable segmentation process is selected according to a priori knowledge of the sample content and speed of operation required. The performance limitations for particular algorithms have been determined to aid the selection process.
Statistical image recognition methods based on linear transformation require a lot of calculation of correlation between subimages and reference patterns of the specified objects to be detected. Image screening provides an effective preprocessing method for lowering the calculation load and improving recognition accuracy. It selects candidate subimages that are similar to the detecting objects in images and rejects the remainders using spatial filters which are based on linear transformation. We have already investigated the spatial filters that are based on 2D projection pursuit (PP). PP requires more heavy calculation load than the principal components analysis (PCA). We, therefore, compare spatial filters based on two kinds of linear transformation algorithms, the PP and PCA, in terms of recognition accuracy and efficiency. Experiments are made for two object detection tasks: eye- and mouth-area detection in face images and text-area detection in document images. The results show that PCA-based image screening is superior to PP-based one for the eye- and mouth-area detection. PCA also achieves higher recognition rate (75%) than PP for the eye- and mouth-area detection, while PP offers equal performance in text-area detection. The results suggest that PCA is totally superior to PP in image screening.
In this paper a neural network to perform surface defects detection in process control and automated inspection is described. Since a lot ofmaterials and defects to be inspected are rich in textural information, we propose to perform automated inspection using a texture classifier. In particular, in this work, we will ' deal with oriented textures, that are characterized by a dominant local orientation everywhere, varying locally and without a primitive element. We analyze an oriented texture image by representing it in a vector space: to each point (,y)is associated a 2D vector with the direction recovered from the dominant local orientation and the module proportional to the coherence (the degree of anisotropy of the vector image). The vector field is analyzed by projecting it on a set of linearly dependent vectors, which may or may not completely span the vector space: we find the optimal projections onto each one by satisfying global optimization criteria using a least-square-error technique implemented on an adaptive neural network. The coefficients of the projection in the basis vector are the texture parameters by means of which the texture classification is performed. Keywords: industrial inspection, oriented texture analysis, orientation field, vector space, parameter space, neural network.
We present a method designed to solve the problem of automatic color grading for industrial inspection of textured ceramic tiles. We discuss problems we were confronted with, like the temporal and spatial variation of the illumination, and the ways we dealt with them. Then, we present results of correctly grading a series of textured ceramic tiles, the differences of which were at the threshold of the human perception.
Parametric estimation is achieved for the discrete 2D Boolean model by applying maximum- likelihood estimation on linear samples. Under certain conditions, a 2D Boolean model induces a 1D Boolean model so that the likelihood function of a 1D observation is expressed in terms of the parameters of the 2D inducing model, thereby enabling maximum-likelihood estimation to be performed on the 2D model using linear samples.
In this paper we develop digital image processing techniques to describe the microstructure of fibrous polymeric materials characterized by random texture patterns and to predict their performance. The proposed techniques are discussed in the content of a fibrous non-woven material, obtained by melt blowing polypropylene resin and used for air filtration. The material microstructure is described by measurements of fiber orientation and estimates of pore area and perimeter. The latter involves using either detailed image analysis or a statistical model, in this case Poissonian line tessellation. The first approach is appropriate for laboratory studies while the statistical model lends itself to on-line quality assurance applications.
Cracks occurred in aircraft engine parts have to be detected as early as possible to prevent engine failure. Fluorescent Penetrant Inspection (FPI), that applies fluorescent materials on metallic surfaces for flaw detection, is a generally accepted technology for nondestructive inspection of surface cracks. The major problem with application of FPI technology is the costly false alarms caused by non-crack fluorescence indications (noise), especially when inspecting used engine parts. A novel crack-detection system for automatic FPI of engine parts using image processing and pattern recognition theories is presented. A strong noise reduction capability and a small number of reliable features for pattern recognition are the two primary characteristics of the system, which contains three major modules: noise-reduction and preclassifier module, feature extraction module, and pattern recognition module including four pattern classifiers. An image synthesizing technique is developed to simulate real-world situations by combining the segmented fluorescence images of man-made cracks with the noisy background of fluorescent images captured from actual used parts. The designed system can eliminate over 80% of noise while retain 94% of crack indication. The total error rate using Fisher's linear classifier is less than 3%, with only 4% of crack misclassification.
High integrity castings require surfaces free from defects to reduce, if not eliminate, vulnerability to component failure from such as physical or thermal fatigue or corrosion attack. Previous studies have shown that defects on casting surfaces can be optically enhanced from the surrounding randomly textured surface by liquid penetrants, magnetic particle and other methods. However, very little has been reported on recognition and classification of the defects. The basic problem is one of shape recognition and classification, where the shape can vary in size and orientation as well as in actual shape generally within an envelope that classifies it as a particular defect. The initial work done towards this has focused on recognizing and classifying standard shapes such as the circle, square, rectangle and triangle. Various approaches were tried and this led eventually to a series of fuzzy logic based algorithms from which very good results were obtained. From this work fuzzy logic memberships were generated for the detection of defects found on casting surfaces. Simulated model shapes of such as the quench crack, mechanical crack and hole have been used to test the generated algorithm and the results for recognition and classification are very encouraging.
We present a method for automatic detection and removal of defects on flash X-ray images. In our experiments, we obtain about ten images representing the radiographic projection of the same object. Defects appear on some images and limit the measurement accuracy. Our aim is to assess the radiant image (image before detection) by combining the relevant information on the different radiographs. First, we isolate the defects from the object projection by the point by point difference between two images. Then, a multiscale filtering is performed on the difference images to extract the defects according to their spatial extent. Next, each filtered image is segmented by a fuzzy clustering method which takes into account inaccurate edges of the defects. The resultant fuzzy images indicate the defect importance on the difference images by a degree between 0 and 1. Afterwards, the fuzzy images obtained are aggregated to retrieve the fuzzy relevance degrees of the measurement on each original image. Finally, the defects are removed by a fuzzy combination of the images according to their relevance degrees. On the resultant images, the defects are well suppressed thanks to the selection of the reliable information.
Lace is particularly difficult to inspect using machine vision since it comprises a fine and complex pattern of threads which must be verified, on line and in real time. Small distortions in the pattern are unavoidable. This paper describes instrumentation for inspecting lace actually on the knitting machine. A CCD linescan camera synchronized to machine motions grabs an image of the lace. Differences between this lace image and a perfect prototype image are detected by comparison methods, thresholding techniques, and finally, a neural network (to distinguish real defects from false alarms). Though produced originally in a laboratory on SUN Sparc work-stations, the processing has subsequently been implemented on a 50 Mhz 486 PC-look-alike. Successful operation has been demonstrated in a factory, but over a restricted width. Full width coverage awaits provision of faster processing.
Much progress has been made in using computer vision to automate the process of lace scalloping. Because the material is flexible, dealing with stretch due to mechanical feed produces a challenge. The vision system has to work with many different patterns and sizes of lace as well as tolerating misalignment. A Fuzzy Reasoning Rule-based technique is employed in order to overcome the problems of material flexibility. Several experiments have been carried out using lace patterns of varying complexity. All cutting paths across the patterns were successfully found. Experimental results indicate that this method can correctly detect the river path in different lace patterns, and cope with lace stretch as well as distortion.
Machine vision systems are increasingly being applied to the apparel industry for the inspection of web fabrics', complete garments2, lace3 and others. This paper describes a system component (delineation) from an automatic web fabric inspection system developed at De Montfort University, which is described more fully elsewhere1. A typical approach to detecting defects is to apply a threshold to the electronic signal generated by a CCD camera scanning the surface. The surface structure of the material introduces a noise component which tends to mask the message signals arising due to defects, making them hard to detect. The presence of a defect causes the signal to temporarily deviate from its mean noise position, and a dual threshold can be used to detect these deviations and generate message triggers, as shown in figure 1. The position of the thresholds can be determined by trading off the probability of false alarm with the probability of a correction detection. To detect defects of low contrast, the thresholds must be placed close to the noise mean, but this dramatically increases the probability of a false alarm trigger arsing. The ternary signal then undergoes a stage of filtering to remove isolated (and hence probably noise) triggers.
A system is described for production-line mail packaging inspection and validation. This system utilizes optical character recognition to achieve real-time performance. A computationally efficient technique is employed for binary image noise reduction. Another effective method has been developed for fast and concise text segmentation. An improved character feature extraction technique has been adopted for robust numeral recognition. A data structure is implemented to enable the system to simultaneously track and process multiple targets. High recognition rate and real-time performance is demonstrated by a prototype system.
This paper presents a general process for an automatic reading of analog instruments. The calibration of analog watermeters serves for a demonstration of this algorithm. Reading a measuring instrument means detecting the position of scales and pointers in the intensity image in order to determine the value displayed by the measuring instrument. The reading process involves the detection of its boundaries, scales, pointers and lettering elements by using pattern recognition algorithms and a-priori information about the geometry and the elements of the meter. First, a validation for the correct type has to be done by checking the actual, detected elements with the elements defined in the type-description. Following the determination of the indicated measurement value, the intensity image of the watermeter is compressed and stored to ensure a given quality standard by protocolling each measurement step. The calibration process consists of five reading phases. A series of watermeters of the same type is calibrated in one run with water quality measurement. The outlined automated approach to the problem is faster, more reliable, stable and cheaper than the manual reading and checking of the meters performed now. Results concerning time, accuracy and reliability are given at the end of the paper.
A machine vision inspection system is designed and built for automatic inspection at the end of automobile gauge panel production line. The inspection items on the gauge panel are pointing errors on all scales of 5 indicators and possible damage or missing assembled warning lights and light bulbs for indicators. Image acquisition camera is set to have a small field of view, a CNC system is established to drive the camera focusing on any target on the gauge panel. The position of the camera is close-loop controlled by a image character feedback control strategy. Automatic calibration is performed by using a stochastic adaptive control scheme. A two-CPU computer system is established to assure real time image processing and CNC control as well as test signal source management working in parallel way. Precision test signal source for speedometer, petrol gauge, oil pressure indicator, water-thermometer and rheometer are designed and made integrated under computer management and control. Each scale and pointer on the gauge panel has a set of image processing parameters, therefore a learning sequence method is designed to reduce programming load and increase flexibility which allows quick adaptation to various products inspection.