基于深度学习的遥感图像目标检测方法研究

Research on remote sensing image target detection method based on deep learning

针对目标小、分布集中的目标遥感图像检测准确率低、性能差等问题,提出了具有自学习能力的深度学习遥感图像目标检测方法,引入注意力机制进行多尺度多分辨率的特征自学习及融合挖掘,改进多尺度单阶段网络目标检测模型SSD的多层特征模块,在模型训练阶段采用改进的感知损失函数避免样本差异过大造成的不均衡问题。实验阶段,对原始SSD算法、FPN算法进行比对实验,文中提出算法的准确率提升在6%~8.6%,速度上也有了明显的改观。结果显示,文中提出的算法对于目标检测具有更好的检测效果,有一定的研究价值。

Based on the problems of low accuracy and poor performance of remote sensing image detection with small target and centralized distribution,a deep learning remote sensing image target detection method with self-learning ability is proposed.The attention mechanism is introduced to carry out multi-scale and multi-resolution feature self-learning and fusion mining,and the multi-layer feature module of multi-scale single-stage network target detection model SSD is improved,In the model training stage,the original SSD algorithm and FPN algorithm are being compared,indicating that the accuracy of the algorithm proposed in this paper is improved between 6%and 8.6%,and the speed has also been significantly improved.The results show that the proposed algorithm has better detection effect for target detection and has certain research value.