An image segmentation system which uses a connectionist classifier architecture as a central building block is described in this paper. The complete system, which consists of a feature extraction module and the connectionist classifier module, has been designed and implemented in digital VLSl; system architectural aspects as well as the procedure of adaptation of the system to different segmentation problems are discussed. The performance of the segmentation system on real world problems is demonstrated using scenes from industrial inspection, texture recognition, and combustion chamber research tasks.

A connectionist classifier architecture that is used in an image segmentation scheme based upon a decision analysis paradigm is described. An appropriate architecture for a connectionist classifier is discussed, and its relationship to other image segmentation architectures following the same paradigm is shown. The performance of a complete system, including pixel descriptor extraction and connectionist classifier, is demonstrated using scenes from industrial inspection and combustion chamber research tasks.>

The architectural and implementation issues of a prototype of a low-level image segmentation architecture (LISA) are described. LISA performs real-time gray-level image segmentation. A decision-theoretic pattern recognition approach divided into feature extraction and decision analysis parts is used. The feature extraction part is based on extracting local and global descriptions of the image pixels. In the decision analysis part, the pixels are classified into one of the user-selected classes based on the extracted features. One of the main properties of the approach is that the feature extraction and decision analysis are performed in real time. The status of the hardware and software development of the LISA prototype for a variety of industrial inspection applications is described. The LISA architecture is being implemented using state-of-the-art VLSI technology.>

A method of feature selection is presented that has linear computational complexity, and ways are shown how to use it when the type of the probability density function is unknown. There is no claim that the procedures for nonparametric probability density function estimation are applicable to any thinkable distribution, but the lower bound for the classifier performance estimation, makes the presented measure applicable in most practical cases. Simulations with synthetic test data as well as references to applications with real-world data demonstrate the applicability of the measure discussed.>

Machine vision methods are presented for the analysis of solder balls in integrated circuits. The algorithms are founded on counter fitting using a multiparameter Hough transform and on polynomial-classifier-based pattern recognition. The first method is used to show the complexity of the inspection problem, especially in the presence of high-precision requirements. In this connection, it is shown that subpixel accuracy is not obtainable even under the assumption of a perfect camera system which determines the resolution necessary for the measurement of a given maximum-volume distortion. The second method is carried out by computing a large number of features on the original image after individual solder balls are segmented by a projection technique. This approach can be considered as a control-free image segmentation paradigm, since it does not rely on properly sequencing several image-analysis modules. Further experimentation with a large pool of defective solder balls is necessary to confirm the applicability of these machine vision algorithms to a real-world manufacturing inspection systems. A general image-segmentation architecture is proposed, which enables the computation of the necessary low-level image features as well as pixel classification at video-rate speed.>

In this paper, we tackle the problem of recovering an image from its Fourier transform phase quantized to 1 bit, or, equivalently, from the zero crossings of the real part of the Fourier transform. We first present new theoretical results that set an algebraic condition under which real zero crossings uniquely specify a band-limited image. We then show, however, through a large-scale set of experiments, that sampling in the frequency domain presents a major obstacle to good reconstruction resuits due to the information loss produced by the approximated knowledge of the zero crossing locations. We finally show that, by using a "coherent" image model in which the image is complex and the spatial-domain phase is random and highly uncorrelated, we can significantly reduce the effect of this information loss and improve the quality of image reconstruction.

We address the problem of segmenting and analyzing images of solder balls in integrated circuits. Two methods are explored, one based on projections computed along curved contours derived from the geometry of the problem, and the other using a supervised, feature-based polynomial classification method. High accuracy is required for the machine vision solution to be successful, and we have included some considerations concerning sub-pixel accuracy for the computation of the values of parameters determining the two-dimensional solder ball shapes. The classification technique has performed with satisfactory results on available data.