基于时频域深度网络的海面小目标特征检测

为了提升海杂波背景下小目标探测性能，本文提出一种基于时频域深度网络的特征检测方法。首先，将观测向量转换为归一化时频图（Normalized Time-Frequency Graph, NTFG），实现海杂波抑制。在时频域，建立海杂波、含正多普勒偏移目标回波、含负多普勒偏移目标回波的三分类问题，精细化目标落在主杂波带内外的不同特性。其次，引入Inception-ResNet V2深度网络作为特征提取器，自主学习不同类别在NTFG上的深层差异性，并将差异性浓缩为一个2D特征向量。然后，在2D特征空间中，设计具有引导的三次样条曲线，获得虚警可控的判决区域，实现异常检测。最后，IPIX实测数据验证了所提算法的性能优势，能深入挖掘时频域的特性。

Feature Detection of Small Sea-Surface Target via Deep Network in Time-Frequency Domain

In order to improve the detection performance of small targets in sea clutter, a feature detection method based on deep network in time-frequency domain is proposed in this paper. Firstly, observation vector is transformed into the normalized time-frequency graph (NTFG) to suppress sea clutter. In time-frequency domain, a triclassification problem of the sea clutter, the target echoes with positive Doppler shift and the echoes with negative Doppler shift is established to elaborate the different characteristics of targets falling inside and outside the main clutter. Secondly, the Inception-ResNet V2 deep network is introduced as feature extractor to independently learn the deep differences from different categories on NTFG, and these differences are condensed into a 2D feature vector. Third, in 2D feature space, the cubic spline curve with guidance is designed to obtain the decision region with controllable false alarm and realize anomaly detection. Finally, it is verified by IPIX measured data that the proposed algorithm can attain good performance and can dig into the characteristics of time-frequency domain.