DBU-Net Based Robust Target Detection for Multi-Frame Track-Before-Detect Method

Abstract

The multi-frame track-before-detect (MF-TBD) method has excellent detection performance for weak targets. However, the statistical characteristics of the merit function after accumulation of multiple consecutive frames are complex, and the setting of the constant false alarm threshold is difficult, especially when the background statistical characteristics are unknown and nonhomogeneous. This paper considers the robust target detection method for MF-TBD. The weak target detection in the merit function domain plane is modeled as binary classification of pixels on the plane. Due to the motivation of classifying pixel points, the U-Net network is selected. Then we improve U-Net into a novel DBU-Net network structure, and train DBU-Net through different merit function domain sample sets. The DBU- Net can effectively detect target in the merit function domain, although the background statistics are unknown and nonhomogeneous. The simulation results demonstrate the superiority and robustness of the detection performance of the method.