Mean-shift based object detection and clustering from high resolution remote sensing imagery

Object detection from remote sensing images has inherent difficulties due to cluttered backgrounds and noisy regions from the urban area in high resolution images. Detection of objects with regular geometry, such as circles from an image uses strict feature based detection. Using region based segmentation techniques such as K-Means has the inherent disadvantage of knowing the number of classes apriori. Contour based techniques such as Active contour models, sometimes used in remote sensing also has the problem of knowing the approximate location of the region and also the noise will hinder its performance. A template based approach is not scale and rotation invariant with different resolutions and using multiple templates is not a feasible solution. This paper proposes a methodology for object detection based on mean shift segmentation and non-parametric clustering. Mean shift is a non-parametric segmentation technique, which in its inherent nature is able to segment regions according to the desirable properties like spatial and spectral radiance of the object. A prior knowledge about the shape of the object is used to extract the desire object. A hierarchical clustering method is adopted to cluster the objects having similar shape and spatial features. The proposed methodology is applied on high resolution EO images to extract circular objects. The methodology found to be better and robust even in the cluttered and noisy background. The results are also evaluated using different evaluation measures.