深度学习遥感影像近岸舰船识别方法  

Inshore Ship Recognition in Optical Remote Sensing Imagery Using Deep Learning

作　　者：王昌安 田金文[1] 张强[2] 张英辉[2] WANG Chang’an;TIAN Jinwen;ZHANG Qiang;ZHANG Yinghui(School of Artificial Intelligence and Automation,Huazhong University of Science and Technology,Wuhan 430074,China;Beijing Institute of Spacecraft Systems Engineering,Beijing 100086,China)

机构地区：[1]华中科技大学人工智能与自动化学院,武汉430074 [2]北京空间飞行器总体设计部,北京100086

出　　处：《遥感信息》2020年第2期51-58,共8页Remote Sensing Information

基　　金：国家自然科学基金项目(61273279)。

摘　　要：针对复杂背景近岸舰船检测与细粒度识别难题,提出了一种基于深度学习的新型端到端目标识别框架,可有效检测与识别任意方向的舰船目标。针对舰船目标短边尺度较小问题,提出了角度致密化的预设框设置方法,提高了候选区域生成时的召回率;采用改进方位敏感型区域插值池化,减少了坐标量化误差,实现了舰船局部区域特征的精确建模;利用注意力机制下的全局与局部特征区域级融合方法,提升了区域特征的类别判别能力,解决了细粒度舰船识别难题;针对舰船样本稀缺性问题,使用迁移学习提升了模型性能。构建了一个含有25类近岸舰船目标的细粒度数据集,与传统学习模型相比召回率提高2%,平均识别精度提高3%,对复杂背景下目标识别具有重要实用价值。To better solve the problem of inshore ship location and fine-grained recognition under clutter background,an endto-end oriented ship detection and recognition framework powered by deep learning is proposed.To deal with the extreme small scale of ship objects,a novel anchors angle density strategy is introduced to recall smaller ship regions in the proposal stage.To avoid any quantization of the region boundaries,the rotate position sensitive region of interest align pooling module is proposed to extract more accurate local ship features.An attention mechanism based global and local region feature fusion method is adopted to enhance the feature discriminating power of ship regions for fine-grained classification.Moreover,transfer learning is also used to further improve the performance of the model under limited samples.Ablation experiments are conducted on the fine-grained inshore ship dataset with 25 classes,and the final model leads to 2% recall rate and 3% average precision rate improvement compared to the traditional model,which has high practical application value for inshore ship detection and recognition under clutter background.

关 键 词：近岸舰船检测 细粒度分类 深度学习 端到端学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]