基于时频图自主学习的海面小目标检测

多维特征检测方法已成功运用于海面小目标探测中。针对人工特征提取的局限性,将检测问题转变为两分类问题,提出了一种基于时频图深度学习的目标检测方法。首先,将一维观测回波变换到二维时频域,并通过归一化时频图进行白化预处理。其次,建立半仿真含目标回波数据库,解决两类训练样本非均衡问题。然后,搭建迁移学习模型自主学习时频图特性,具有深度网络结构和减小训练代价的优势。最后,将两分类的概率值作为统计量,获得虚警可控的判决区域。基于IPIX实测数据实验结果表明:所提的检测器能深入挖掘目标和杂波的差异性,低信杂比下仍能有效提升海面小目标的探测能力。

Sea-surface small target detection based on autonomic learning of time-frequency graph

Multi-dimensional feature detection method has been successfully applied to sea-surface small target detection. Due to the limitation of artificial feature extraction, detection problem is converted into two classification problems and the target detection method based on deep learning of time-frequency (TF) graph is proposed. Fristly, one-dimensional observation echo is transformed into two-dimensional TF domain, and the whitening pre-treating is carried out by normalized TF graph. Second, a semi-simulated database with target echo is established to solve the problem of unbalanced two training samples. Then, transfer learning model is built to independently study the properties of TF graph, which has the advantages of deep network structure and reducing training cost. Finally, the probabilities of two class serve as statistics to obtain the decision region with controllable false alarm rate. Experimental results by IPIX datasets show that the proposed detector can deeply mine the differences between target and clutter, which can effectively improve the detection capability of sea-surface small targets at low signal-to-clutter ratio case.