Scale invariant SURF detector and automatic clustering segmentation for infrared small targets detection
Haiying Zhang,Jiaojiao Bai,Zhengjie Li,Yan Liu,Kunhong Liu
DOI: https://doi.org/10.1016/j.infrared.2017.04.001
2017-06-01
Abstract:The detection and discrimination of infrared small dim targets is a challenge in automatic target recognition (ATR), because there is no salient information of size, shape and texture. Many researchers focus on mining more discriminative information of targets in temporal-spatial. However, such information may not be available with the change of imaging environments, and the targets size and intensity keep changing in different imaging distance. So in this paper, we propose a novel research scheme using density-based clustering and backtracking strategy. In this scheme, the speeded up robust feature (SURF) detector is applied to capture candidate targets in single frame at first. And then, these points are mapped into one frame, so that target traces form a local aggregation pattern. In order to isolate the targets from noises, a newly proposed density-based clustering algorithm, fast search and find of density peak (FSFDP for short), is employed to cluster targets by the spatial intensive distribution. Two important factors of the algorithm, percent and γ, are exploited fully to determine the clustering scale automatically, so as to extract the trace with highest clutter suppression ratio. And at the final step, a backtracking algorithm is designed to detect and discriminate target trace as well as to eliminate clutter. The consistence and continuity of the short-time target trajectory in temporal-spatial is incorporated into the bounding function to speed up the pruning. Compared with several state-of-arts methods, our algorithm is more effective for the dim targets with lower signal-to clutter ratio (SCR). Furthermore, it avoids constructing the candidate target trajectory searching space, so its time complexity is limited to a polynomial level. The extensive experimental results show that it has superior performance in probability of detection (Pd) and false alarm suppressing rate aiming at variety of complex backgrounds.
optics,physics, applied,instruments & instrumentation