作者: Yongtao Yu , Haiyan Guan , Zheng Ji
DOI: 10.1109/LGRS.2015.2432135
关键词:
摘要: This letter presents a rotation-invariant method for detecting geospatial objects from high-resolution satellite images. First, superpixel segmentation strategy is proposed to generate meaningful and nonredundant patches. Second, multilayer deep feature generation model developed high-level representations of patches using learning techniques. Third, set multiscale Hough forests with embedded patch orientations constructed cast votes estimating object centroids. Quantitative evaluations on the images collected Google Earth service show that an average completeness, correctness, quality, $F_1$ - measure values 0.958, 0.969, 0.929, 0.963, respectively, are obtained. Comparative studies three existing methods demonstrate superior performance in accurately correctly arbitrarily oriented varying sizes.