Detection of noisy rectangular targets in high spatial resolution images

Abstract

ABSTRACTHuman-related targets in high-spatial-resolution images often present rectangular shapes, such as most buildings, farmland, etc. However, their edges often encounter bending or missing parts as well as redundancy, making them difficult to detect by current breaking-assembling strategies that initially detect line segments and then gather them into potential rectangles. We propose a rectangle-detection method that uses a whole-object detection strategy that detects rectangles as integrated units to avoid these problems. Comparison experiments with a state-of-the-art method on both synthesized and high-spatial-resolution images indicate that our method performs well in detecting rectangles both in accuracy and time-cost aspects, especially when the edges are noisy.