Computational Structures for Extracting Edge Features from Digital Images for Real-Time Control Applications

This thesis investigates extracting edge features from digital images for real-time control applications. A sequence of image operations is presented for extracting edge features from a digital image. This sequence emphasizes intensity edges, isolates them from the rest of the image and then converts them into line data structures for further analysis by a higher-level vision processor. This way, an image of 500 x 500 pixels is reduced to a few hundred data elements. In order to extract edge features from digital images fast enough for real-time use, some consideration has been given to the computational structures for performing the edge-extraction operations. Past attempts at constructing parallel processing arrays to speed up image processing are summarized. From examining these past efforts, it is concluded that array processing is unsuitable for real-time image processing. One of the reasons for rejecting the use of array processors is the difficulty in constructing arrays large enough to cover reasonable-size images. A study reveals that this difficulty will remain even when ultimate LSI gate densities are used in constructing the arrays. Specialized processors are found to be more suitable for real-time image processing. From examining the computational characteristics of the edge-extraction algorithms, it was found that all but one of them could be suitably executed by pipeline processors. The exception is the chain-coding process, which is more suitably performed by a random-accessed structure. To demonstrate methodology and feasibility, a pipeline unit to perform the edge-enhancement operation in real-time has been designed and constructed using MOS LSI technology. A vision system for real-time control purposes must be able to track moving objects. A brief scenario for using the edge features extracted from a sequence of images to track a moving object is presented in the last chapter of this thesis.