Q+A CENTER

Automated image analysis has long been the driver behind industrial machine vision systems. The technology is mature, and almost mainstream from a challenge standpoint. Still, there are some notable issues that can arise, such as the need to process and analyze a flood of images in real time. Many of the objects being imaged, while uniform in nature, can also be highly complex or pose a variety of imaging challenges due to color issues or environmental factors. With the proper knowledge, hardware, software and tweaking such issues are usually overcome.

Beyond industrial machine vision, similar needs are now being addressed in a broad range of fields. Often, the basic algorithms are similar, at least at the starting point. But, the challenge in applying those algorithms can increase by an order of magnitude to meet the needs of military and security professionals, biological researchers, medical professionals and those who rely on geological and environmental information—to name but a few.

One of the most challenging of these areas is remote sensing, where the goal is to call out details from imagery generated by airborne or orbital platforms using a range of sensors. Typical challenges include the limited control over illumination; an ever-changing background clutter; variable imaging optical perspectives; typically very large images; hyperspectral images with hundreds of channels; variable weather conditions (clouds, snow, etc.) and literally an unlimited number of objects that could be modeled for extraction and that may vary in greatly in appearance while serving the exact same purpose.

"The basic tasks are more or less the same compared to industrial machine vision—you use edge detection or classification or segmentation," said Dr. Christian Wiedemann, a software engineer for MVTec Software GmbH, (Munich, Ger.). MVTec is a leading software developer for machine vision applications, with its HALCON product being used for a range of tasks including basic remote sensing analysis. "But, you have to use more knowledge about the contents of the scene if you want to interpret natural scenes. For instance, identifying context objects which are near the object you want to extract. You can use these objects, if you can find them, to get some information about the specific object you are interested in. However, it is relatively complex to model the relationships between these different objects. What is easier in remote sensing tasks (relative to industrial machine vision), is that you have a lot more time to process the images. You do not have to do this in real time so you can have an hour or even more to process the image."

The challenges are such that most civilian-grade systems are currently limited to less granular applications like land-use classification, where forest, farm and urban terrain can be identified and quantified. Beyond that, identifying and extracting even common elements like roads can pose complex challenges. "Imagine you want to extract roads from an image, and the road goes through a forest so you cannot see the road in those areas," said Wiedemann. "You have to conclude that we have a road found in the open areas on both sides of the forest so the road must go through the forest. So you cannot extract everything from the image."