Architectures for Multidimensional Low- and Intermediate Level Image Processing (Invited)

In the last decade many architectures for low-level image 2Image processing tasks. processing were developed. But simultaneously the image processing tasks were growing and changing, resulting in a As we look at tasks that are performed in image processing different set of basic operations than was needed ten years nowadays a few profiles can be sketched, without ago. An attempt is made to classify the existing and pretending to beexhaustive: proposed architectures over the past ten years followed by the results of a theoretical comparison on three of the main 2.1 Industrial inspection and robotics: groups of architectures, the pipelines, the processor arrays and the linear processor arrays. One of the conclusions is Industrial inspection and robotics is nowadays a main that recursive neighbourhood operations appear to be an application field of the classical 2 0 image processing efficient way to implement global operations and object sequence. operations. But is there support from the architectures? Image acquisition in this field involves both CCD video Another conclusion is that queue or bucket updating is an cameras and line-scan cameras. Gray-value filtering is extremely efficient way to process pixels, especially when avoided by solving the illuinination problem and many elements have to be processed as in 3D image segmentation is typically done by thresholding followed by processing. With the amount of elements to process in a 3D post-processing mostly in the form of morphologic image, the idea of 'one pixel or voxel one processing operations. Measurements are often length of lines, element' has to be abandoned as hardly reachable with the sometimes surface, perimeter and circularity. Pattern current technology. 'Processing only the pixels that change' recognition is mostly done feature based. In this field most points in the direction of processing pixels or voxels only of the commercial real-time image processing systems find on fronts. These kind of operations, such as labelling, their application. E.g. inspection by placement of SMD skeletonization, distance transform and region growing are devices on printed circuit boards, inspection in robot characterized by the name: 'Wave Front Processing'. The assembly sequences. An example of an an advanced support from architectures needed for this type of processing sequence in this field would be the recognition processing is indicated. of 3D objects from a 2D image based on a learning set of images (Bart et a1.1990): Sobel edge detection, isodata