New Methods for Horizon Line Detection in Infrared and Visible Sea Images
Ilan Lipschutz, Evgeny Gershikov, Benjamin Milgrom · 2013
In this work we propose methods for horizon line detection and target marking in marine images captured by either infra-red (IR) or visible light cameras. A common method for horizon line detection is based on edge detection and the Hough transform. This method suffers from serious drawbacks when the horizon is not a clear enough straight line or there are other straight lines present in the image. We improve the algorithm performance by proposing a pre-processing stage eliminating some of the false detections. We also propose a new method for horizon line detection. The new algorithm is based on the idea of segmenting the image into two regions (sky and sea) by comparing the regional probability distribution functions (PDFs) or histograms. The line maximizing a statistical distance between these PDFs is chosen. This method is highly complex and can be further improved. Thus, we proceed to introduce two new methods combining the basic edge detection and Hough transform (EDHT) method to detect several candidate lines in the image and a statistical criterion to find the optimal line among the candidates. The chosen criterion can be based on regional covariances or the distance between PDFs as described above. We show that choosing the first option provides the best performance among the methods tested, at least with respect to the angular error, and yields a low complexity fast algorithm. We compare all the examined methods not only quantitatively by their average accuracy and relative speed, but also visually for several test images. Our conclusion is that all the methods introduced in this work can be beneficial for automatic processing of marine images for the purpose of tracking, navigation, target recognition and other applications.