A knowledge-based technique for robust recognition of bridges in FLIR imagery

In this paper a knowledge-based method for bridge recognition in forward-looking infrared (FLIR) images is proposed. Conventional knowledge-based methods usually employ recursive scanning, clustering of regions and search for coarse water bodies. Our method does not incur these overheads and comprises of three stages. Firstly, edges in the image are detected using a Harris detector, where edges present in highly textured areas are discarded. Secondly, Hough transform (HT) is employed to detect a fixed number of image lines in the bridge's expected orientation. The detected lines are clustered based on different intuitive distance measures and only a few lines are retained for further processing. During the third stage, crosswise neighborhood of the retained lines is examined. The final bridge line is detected based on self-resemblance, non-cornerness, bridge polarity and the presence of piers in the neighborhood. Experiments on real FLIR dataset show that our method is robust to complex backgrounds, noise, low contrast between a bridge and its surrounding area, small bridge-size and bridge orientations.