Tadashi Fukushima

An image signal processor

An image processor using CMOS technology, including about 15,000 gates, will be discussed, which performs most two-dimensional basic processing for 8b gray images. A 4×4 spatial convolution onto a 256×256 picture can be carried out in 10.9ms with four image processors.

A survey of image processing LSIs in Japan

The author reviews image processing LSI ICs which were developed in Japan during the 1980s. Forty devices are covered and classified into five categories: the fully parallel processor, the partially parallel processor, the digital signal processor specialized for image processing, the functional processor, and the neutral network processor.<<ETX>>

An implementation method of multi-viewing on an engineering workstation

We developed the multi-viewing functionality optimized from a viewpoint of CAD users on an engineering workstation. Multi-viewing allows an application program to display a three-dimensional model with multiple scales and angles. Its functionality helps users to grasp a three-dimensional model shape easily. On the other hand, multi-viewing increases the number of graphical elements to be drawn and degrades picture updating speed. One implementation approach of the function is following the PHIGS (Programmers Hierarchical Interactive Graphics System) viewing concept. But it may cause all views to be updated even when a viewing parameter of only one view changes. This makes a CAD user irritated since it takes a long time. To avoid such redundant drawing, the system has to preserve the functionality of updating individual view. So we enhanced the PHIGSs multi-viewing functionality in order to change a picture on an individual view. Two significant functions are realized in our system to be able to change a picture on an individual view. One is of a view updating control, where any view can be updated selectively even if it is overlaid by other views or not. The other is of a view overlapping control, where a view is displayed as if it is overlaid by other views when two or more views are overlapped. First of all, we separates viewing control from modelling management. Modelling is sticked with a data structure, while viewing is attained with viewing transformation parameters building up a view management table. These parameters are to be set before structure traversal. The structure traversal is done depending upon which views parameter changes. That is, the only graphical data which are concerned with a view to be updated are selectively traversed. In addition, we realized view overlapping control by supporting the view output priority which determines whether the view overlays other views. Actual view output area is calculated by referring the output priorities as well as the position of views. Then, an application can redraw the picture only on the actual view output area. As a result, this system realized a overlapped multi-viewing functionality which can draw only a picture selectively related with a particular view and can draw only area of which an application intends to update. This is very useful in a CAD application since the users can quickly change a picture on a particular view in the case of picture rotation and zoom which are frequently used in a CAD operation.

Construction methods and performance evaluation for image processing systems with ISP LSIs

We have developed an LSI, an Image Signal Processor (ISP), especially suited to process image signals of either binary or gray scale. The ISP belongs to the category of partially parallel processing devices and possesses four processor elements (PEs) which operate in parallel. There are two image system construction methods with ISPs. One is the PE adding method, with which four ISPs make up a system which performs a 4 × 4 spatial convolution on a gray-scale image and two ISPs compose one which accomplishes an 8 × 8 template pattern matching on a binary image. These take 10.9 ms onto a 256 × 256 picture. The other method is the PE saving approach, with which four ISPs build up a system which executes an 8 × 8 spatial convolution on a gray-scale image and eight ISPs yield one which carries out a 32 × 32 template pattern matching on a binary image. These need 43.7 ms onto a 256 × 256 picture.