Evaluation of a Class of Segmenters for IR Imagery (Evaluation d'une Classe de Segmenteurs pour Images IR),
DEFENCE RESEARCH ESTABLISHMENT VALCARTIER (QUEBEC)
Pagination or Media Count:
This report presents an evaluation of the performance of 6 algorithms dedicated to segmentation of targets in IR imagery. These segmentation algorithms or segmenters are based on the single assumption that the targets display a larger thermal signature than the background. The first 3 segmenters deal with an image in its entirety, whereas the last 3 incorporate the Background Elimination Technique BET, which aims at eliminating wholly or partly the background by levelling it. The segmenters are judged according to a their extraction rate b the fidelity of the segmentation with respect to the geometrical properties of the extracted targets and c the degree of distinctiveness imparted to the extracted targets as opposed to the nontargets. The 3 segmenters relying on BET have a better extraction rate than the other 3 that try to cope with the background simply by partitioning the image. Most segmenters here distort in one way or another the shape of the targets. The two exceptions are segmenter No. 1 Single Intensity Threshold Silhouette Generator or SIT Generator and No. 6 SIT Generator in conjunction with a particular version of BET. The experimental results show that the intensity, contrast and variance features are the most effective in discriminating the targets from the nontargets.
- Infrared Detection and Detectors
- Human Factors Engineering and Man Machine Systems