Tensored Generalized Hough Transform for Object Detection in Remote Sensing Images

To avoid using a large 4D-Hough counting space (HCS) and complex invariant features of generalized Hough transform (GHT) or its extensions when detecting objects in remote sensing image (RSI), a tensored GHT (TGHT) is proposed to extract object contour by simple gradient angle feature in a 2D-HCS using a single training sample. Considering that tensor can record the structure relationship of object contour, tensor representation R-table is constructed to record the contour information of template. For slice centered at each position of RSI, the tensor-space-based voting mechanism is presented to use the tensor that records the contour information of slice to gather votes at the same entry of 2D-HCS. Furthermore, a multiorder binary-tree-based searching method is presented to accelerate voting by searching the index numbers of elements in tensors. In addition, by solving the tensor-space-based optimization problem that is used to determine the candidates objects, the cause of false alarms (FAs) caused by interferences with complex contour and FAs caused by interferences that are partial-similar to objects is revealed, and the matching rate and matching sparsity-based strategies are then proposed to remove these FAs. Using public RSI datasets with different scenes, experimental results demonstrate that TGHT reduces nearly 99% storage requirement compared with GHT for RSI with size exceeding 1000 x 1000 under small time consumption, and outperforms the well-known contour extraction methods and state-of-the-art deep-learning-based methods in terms of precision and recall.