基于多特征融合的SAR图像舰船自学习检测算法

Self-learning Detection Algorithm for Ship Target in SAR Images Based on Multi-feature Fusion

传统的舰船检测方法主要包括恒虚警检测(CFAR)和机器学习类算法,其中CFAR舰船检测容易受噪声影响,检测结果过分依赖参数与海杂波模型的选择,准确率低并且鲁棒性较差。简单的阈值判定方法由于特征单一,对舰船目标描述性较差,机器学习算法需要对已有数据库中舰船数据进行训练,准确度较高但检测周期过长,更新较慢,无法满足现代战争的快速反应、实时更新的需求。提出一种基于多特征融合的自学习算法,对感兴趣目标提取形态学、灰度和轮廓等多种特征,通过对多特征阈值判定方法对相似舰船目标进行检测,可实现对战场突发状况与未知目标快速反应能力的同时保证较高的检测准确率。实验结果表明,提出的检测算法相比传统方法的查全率提高了10%,虚警率降低了4%,并且实现了单幅运行时间的大幅度缩减。

The traditional ship detection methods mainly include Constant False Alarm Rate （ CFAR） and machine learningmethods,the CFAR ship detection is susceptible to noise,the test results are excessively dependent on the choice of parameters and clutter model, so the accuracy rate and robustness of these results are low. The simple threshold determination method has such disadvantages as single feature and poor description of ship target,the machine learning algorithm needs to perform the training for ship data in the existing database,so the accuracy is higher,but the detection period is too long,the update is slow,and it can’ t meet the requirements of modern war for fast response and real-time update. This paper proposes a self-learning method based on multi-feature fusion to extract morphological,gray and contour features from interested objects. By detecting multi-feature thresholds,the similar ship targets can be detected, and the higher detection accuracy rate can be ensured while realizing the fast response for emergency and unknown target. The experiment results show that the proposed detection algorithm can raise the recall ratio by 10%,decrease the false alarm rate by 4%,and shorten significantly the running time of single image,compared with traditional methods.