PC-based machine vision systems will give way to monolithic vision sensors with smaller, cheaper, integrated computing devices that are simpler to implement and to network. Most of the theoretical advances that form the basis of modern industrial machine vision are more than 20 years old. To confirm this assertion, one need only review the contents of Digital Picture Processing, edited by A. Rosenfeld and A. Kak (Academic Press, Vols. 1 and 2) and published in 1982. A machine vision system is simply a computer capable of image capture and processing. Therefore, it’s not surprising that the recent history of machine vision reflects the adaptation of evolving computer technology to commercial image processing as applied to industrial automation. This process has passed through several stages, each heavily influenced by advances in semiconductor and computer software technology. During the 1970s, general-purpose mainframe computers were coupled with image capture and display peripherals. Throughout the 1980s, mainframes gave way to dedicated computers and proprietary hardware designed to accelerate image processing. During the 1990s, machine vision followed the prevailing trend toward standard platform computing. High-performance microprocessors and low-cost personal computers made most special-purpose vision hardware obsolete. Similarly, today’s advances in high-speed networking and mobile computing have positioned the machine vision industry for the next step: integration of the complete machine vision system into a monolithic network appliance. The vision system is becoming the vision sensor, a tightly integrated system that encloses imaging, processing, memory and connectivity functions in a single box. Such sensors have already begun to eliminate the expense and footprints required for a computer workstation, but the next generation of sensors will deliver performance and ease of use at a significantly lower cost. New vision sensors may be smaller, easier to use and less costly than computers, but another measure by which they will succeed is connectivity. How well machine vision systems lend themselves to networking is an important criterion. The cost of bundling a microprocessor into a digital camera is already small, but the increased level of circuit integration will have, among other benefits, an inverse effect on the incremental cost of processing. High-performance processors The physical size and thermal characteristics of image processing electronics had prevented the practical realization of compact vision sensors that could compete with PC-based machine vision systems. However, advances in semiconductor process technology have enabled development of high-performance digital signal processors on smaller dies that operate at lower core voltages. The result is a new generation of processors that can deliver performance up to an order of magnitude better, yet, amazingly, consume little if any additional power. This is an important development because the ability of a vision sensor to passively dissipate heat is extremely limited. With the relaxing of package size and thermal constraints, vision sensors will vie with PCs for the bulk of machine vision applications. The PC/frame-grabber standard prevalent today is burdened by both the expense and the logistics of high-speed image communication. Integrating the processors and image capture functions into a single box reduces the cost of hardware required to transfer images from the sensor to the processing element. Besides being the least expensive alternative, the sensor is a simpler solution. Its architectural simplicity translates directly into ease of use, reliability, maintainability and reusability. Users will begin implementing network-distributed computing techniques to optimize the benefits of vision sensors. Distributed computing will allow the vision sensor to become a self-sufficient network resource, supplying platform-independent user interfaces to clients anywhere on the network, located anywhere in the world. Point-to-point dedicated user-interface devices will gradually become obsolete. In the future, users will be able to configure and monitor their machine vision systems using everything from cell phones to engineering workstations. Beyond the simple client-server structure, vision sensors will enable peer-to-peer cooperation organized to address complex multicamera inspection and alignment tasks. Applied as elements of a larger parallel processor, they will be scalable to match the needs of almost any application. Expect vision sensors to become smaller, faster and easier to integrate. Expect them to provide all of the algorithms, speed and functionality seen in today’s PC-based machine vision systems. Finally, expect the low cost and simplicity of vision sensors to stimulate the expansion of the industrial machine vision market and to help create new market opportunities for fundamental vision technology. Meet the author E. John McGarry is senior vice president and general manager of Cognex’s In-Sight products group in Portland, Ore.