一种基于深度学习的遥感图像目标检测算法

A deep learning based object detection algorithm for remote sensing images

遥感图像分析在国土资源管理、海洋监测等领域有着极为广阔的应用前景。深度学习技术已在图像处理领域取得突破性进展,然而,遥感图像固有的尺寸大、目标小而密集等特点,使得将面向普通图像的深度学习方法用于遥感目标检测普遍存在定位不准确、小目标检测难、大图检测精度差等问题。针对上述难题,提出了一种新型遥感图像目标检测算法DFS。与传统机器学习方法相比,DFS设计了新的维度聚类模块、定制损失函数和滑动窗口分割检测机制。其中,维度聚类模块通过设计聚类机制优化定制先验框,提高定位精度;定制损失函数提高对船只等小目标的检测精度;滑动窗口分割检测解决大图检测精度低的问题。在经典遥感数据集上开展的实验对比表明,与YOLOv2相比,DFS算法的mAP提高了25.6%,小目标检测效率及大图检测效能大幅提高。

Remote sensing image analysis has extremely broad application prospects in the fields such as land and resources management and ocean monitoring.Deep learning technology has made breakthroughs in the field of image processing.However,due to the inherent characteristics such as large size and small and dense objects of remote sensing,the deep learning methods for common images has some problems such as inaccurate positioning,difficult small object detection,and low large image detection accuracy in object detection for remote sensing images.Aiming at the above problems,this paper proposes a new object detection algorithm DFS for remote sensing images.Compared with traditional machine learning methods,DFS optimizes and designs a new dimension clustering module,customized loss function and sliding window segmentation detection mechanism.Among them,dimension clustering module optimizes priori anchors by designing clustering mechanism to improve the positioning accuracy,the customized loss function improves detection accuracy of small objects such as ships,and the sliding window segmentation detection solves the problem of low detection accuracy of large images.The comparative experiments on the classical remote sensing datasets show that,compared with YOLOv2,DFS improves mAP by 25.6%and greatly improves the detection efficiency of small objects and large images.