Imaging systems that operate in the mid-infrared spectrum have a lot of potential for use in military trackers and weapon systems. Edge detection is the primary means from which a target's spatial and trajectory features are derived. Targets are separated from the image background using edge detection in a process called segmentation. Typical edge detection techniques employ thresholding operators to emphasize the areas in an image where intensity changes significantly. Techniques have been developed which utilize advanced mathematical functions to improve the accuracy and versatility of detector performance. Three advanced edge detectors developed by Canny, Marr & Hidreth and Shen & Castan are candidates for future Army systems.

Utilizing actual infrared image data and edge detection performance metrics, a comparison of the Canny, Marr-Hildreth, and Shen-Castan algorithms has been made. Each detection algorithm processed a common set of test images which included various edge sizes in blurred and noisy conditions. Test images were captured using an imaging camera which operation range is in the mid-infrared spectrum. A target was aligned with the camera and a controlled background which filled the camera's field-of-view. The distance between the camera and the target was 3.5 meters. The target temperature varied from 48 to 90 degrees Celcius while the background temperature remained conastant at 39 degrees Celcius. Images were taken with the camera's optics focused and de-focused. Random noise with a normal distribution was injected digitally. Edge detection algorithms coded in C processed the image data. Detector performance was rated on the distance and position of the detected edge relation to the actual edge.

Under all conditions the superior edge detector has been shown to be the Canny. The Shen-Castan edge detector, which is specifically designed for noise, showed a large decline in performance under blurry conditions most likely due to the dectection of multiple edges. The Marr-Hildreth detector, which didn't address noise or multiple edges, proved to be the worst edge detector.

In general there seems to be a trade-off in detector sensitivity and detector performance. As you make your algorithm more sensitive to edges it will increase it's [sic] performance in noise. However, you increase your chance of detecting multiple edges. It is clear that if you don't address issues of noise and multiple edge detection at all you will be extremely susceptible to both. Experimentation has shown the Canny algorithm superior to both the Shen-Castan and Marr-Hildreth in both noisy and blurry conditions. An Army tracking system is required to perform well under all conditions. Therefore it is recommended that the Canny edge detector be strongly considered for future infrared weapon and tracking systems.