遥感影像船舶检测的特征金字塔网络建模方法

船舶作为海上运输载体,其准确检测在海洋环境保护、海上渔业生产管理、海上交通与应急处置及国防安全应用中都具有重要意义和价值。目前基于目标检测网络的遥感船舶检测方法因末层特征分辨率不足和卷积固定的几何结构,导致网络难以适应小尺度且具有随机朝向、形态多变特征的船舶目标,进而限制船舶检测精度。针对该问题,本文提出一种用于遥感影像船舶检测的特征金字塔网络建模方法。首先引入形变卷积/RoI池化模块,以适应朝向和形态多变的船舶目标;其次借鉴在小目标检测中性能出色的特征金字塔网络的建模思想,采用对称式网络和多尺度特征融合的方式进一步融合高级语义和低级空间信息,提升小尺度目标特征分辨率。在40000幅、船舶目标67280余个的遥感影像数据集上的试验结果表明,本文方法能够有效集成形变卷积/RoI池化和多尺度特征融合方法,相较传统CNN船舶检测方法取得明显提升,在准确率、召回率及F1指标上分别达到85.8%、97.9%和91.5%。

A deformable feature pyramid network for ship detection from remote sensing images

As a carrier of maritime transportation, the accurate detection of ships is of great significance and value in marine environmental protection, marine fishery production management, maritime traffic and emergency disposal, and national defense security applications. In recent years, the remote sensing ship detection method based on CNN (convolutional neural network) is facing big challenges to adapt to small-scale ships with random orientation and morphological characteristics due to insufficient resolution of the final layer features and convolution fixed geometry, thus reducing the accuracy of object detection. In order to tackle this problem, a remote sensing ship detection method based on deformable feature pyramid network with multi-scale feature fusion. First, the architecture of feature pyramid network is adopted to detect small-scale ship object by using a bottom-up refinement process and multi-scale feature fusion. Then, by introducing the deformable convolution and RoI (region of interest) pooling module to adapt to the ship object with random orientation and morphological characteristics, the ship detection accuracy is further improved. Experiments on 40 000 remote sensing images and over 67 280 ship objects demonstrate that the proposed method performs better than CNN. The rate of recall, accuracy, and F₁-Score are 85.8%, 97.9% and 91.5%, respectively.