In this paper, we propose a weld inspection expert system which is enforced by an image processing environment. This system is consisted by the “expert system” and the “image processing” modules. This system has mainly two knowledge bases. The inspection knowledge base is given by some official standards such as JIS and the know-how's obtained from the inspection experts. The image processing knowledge base is used for the selection of image processing methods and for the tuning of over 40 parameters. In this system, the expert system and the image processing modules are unified and support each other in order to reduce the computation costs of image processing and also to reduce the mistakes of judgment in the inspection.
A high-speed image processing system has been developed. The system can execute a wide variety of general image processing operations and achieves high total image processing speed from image input through result output. The system contains two types of processor. The first is the PU array, processor unit array, which is the most significant feature of this system. The array is an LSI pipelined processor and contains sixteen processor units. Compared with previous pipeline processors, the PU array executes highly flexible preprocessing and feature extraction. By writing the desired instructions to each unit and setting the appropriate bus connections among processor units and image memories, a flexible pipeline processor network can be configured. The second type is a microprogrammable main processor that is used to both control the PU array and execute complicated processing tasks which the PU array cannot perform. The main processor is specifically designed for image processing applications and therefore obtains higher speed than general purpose microprocessors.
A new technique has been developed for discrimination between flaws and deposited contaminant particles (dusts) on smooth and transparent surfaces such as photomask-blanks and flat-panel-display substrates. This method is based on the measurement of the angular distribution and polarization characteristics of scattered light which is generated by a focused He-Ne laser beam. Various flaws and dusts with sizes of the micron order were investigated using this method. While the ratio of scattering intensity at different angles was more than 0.20 for dusts, it was less than 0.08 for flaws. The polarization ratio of more than 0.1 was measured for dusts as compared with less than 0.06 for flaws. These results show that each way was enough to discriminate between flaws and dusts. Moreover, it was comfirmed that a combination of ways to measure the angular distribution with polarization characteristics improves the efficiency and accuracy of discriminating the two as compared with the individual measurement mentioned above.
A non-contact method to measure 3-D shape of smooth glossy objects is proposed. This method uses a M-array coded light source to determine correspondence between a position on the light source and a point on an image of the surface of the glossy object which reflect light from the light source, and estimates the 3-D shape of the objects from the correspondence by fitting a plane localy. It requires less light source planes and smaller template for deciding the correspondence than conventional methods. The accuracy of measurement is also improved. The validity of the method is confirmed by experiments for plane mirror.
Machine-vision plays an important role in an automatic inspection for printed circuit boards (PCBs). The conventional image comparison methods, such as direct comparison of two images and image expansion-contraction for design-rule inspection, have some difficulties in finding large defects, missing paterns, and a short. To cope with these problems, topological information on PCBs is introduced in this proposed method. This paper is concerned with an optical inspection system, in which topological information on conductors and insulators of PCBs are used for the detection on defects. The principle is mainly based on the topological comparison method which compares the standard graph obtained from the skeltons of the conductor and insulator images of the standard PCB with those of inspection boards. Topological information incoorporates weighted graphs composed of several types of nodes and edges, connections, and their locations. By this method all of the flaws mentioned above are detected together with their positions and the types of defects are allmost discriminated. Experimental results as well as some pertinent problems are also presented.
This paper proposes a new framework of an instrumentation technique, which deals with human sensibility called “Kansei”, for industrial field, and reports two trials based on this framework to extract factors which decide quality of pearls. This framework consists of three stages. In the first stage, factors are extracted by analysing the responses of human experts. In the second stage, the sensing system is tentatively designed based on the factors, that is for example, to be tuned parameters using the values of proportions. In the last stage, the responses of this system are compared with those of human experts, and the differences are feedback to the first and second stages. Then the system gradually approaches human experts. As a result of our trials in the first stage, several important factors have been extracted by using multivariate analysis and the neural networks.
An algorithm for inspecting the dimensions of automotive body panels using a three-dimensional (3-D) vision sensor has been developed, together with an inspecting system based on this algorithm. The slit light of the 3-D vision sensor is projected onto the body panel and the reference block which gives reference coordinates. In order to measure the gap and flushness between the panel and the block, this algorithm estimates the edge of the bent panel by fitting a circle on the image from the sensor. Further the image is improved by controlling both the projection angle and brightness of the slit light, when the disconnected slit light image is obtained. The system measured the gap and flushness with accuracy to ±0.15mm in the experiments and proved effective for the inspection of panel dimensions.
A high-speed and highly reliable inspection method which inspects minute solder joints of high-density-mounted devices has been developed. A human inspector touches a solder joint by a pinset to identify a “floating lead” defect which is like a good joint cosmetically. The developed method applies force to solder joints by air jet externally and induces vibration or shift of a floating lead. The vibration or shift is detected as a change of speckle patterns which are produced at a solder joint by laser illumination. For reconciling with highly sensitive detection of speckle pattern changes and separable detection of each individual joint in a narrow pitch row of solder joints, defocus amounts in x and y directions are set separately by a cylindrical lens. For decreasing dispertion of detected light intensities of individual solder joints, a liquid crystal filter is inserted in a detection light path and the control of light intensities of individual joints is realized. As defect judgement, two functions which evaluate proportions of vibration and shift from a detected speekle waveform are introduced. The result shows that the developed inspection method detects floating lead defects accurately with a very low false alarm rate.
Recognition of characters stamped on the metal is one of the difficult subject and has a wide application in industrial manufacturing. Conventionally, stamped character recognition has been performed using binarization and segmentation. However, the quality of the background is apt to be poor. Consequently, it becomes difficult to binarize the image. In this paper, recognition process without binarization is proposed. Segmentation and recognition are performed simultaniously by exhaustive pattern matching at every location using the Multi-Angled Parallel Matching Method, where multiple directional planes and a plane for lines are combined as an input feature. The effectiveness is proved by experiments for laser-marked alphanumerics.
Since bend and height errors of QFP IC leads cause assembly defects of components loaded on PCBs. A technique to inspect these errors in QFP ICs is indispensable. An inspection method for QFP IC using a 3-D vision sensor and a television camera has been developed. In the inspection method, the height of leads and the distance between leads, that is referred to as the pitch, are measured by the especially designed 3-D vision sensor. The lead length, which is the distance between the tips of confronting leads, is measured by the television camera. The 3-D vision sensor measures the height and pitch by triangulation with a slit light. In the measurement, slit light projection power and position are controlled in order to restrain the effect of solid dip coating on the lead surface. The developed system containing these algorisms inspects a 44pin QFP IC with a measurement accuracy of ±0.02mm and an inspection time of 7 seconds.