李仕佳

摘要：遥感影像目标检测与识别是近些年来国内外研究的难点，利用深度学习的目标识别技术，直接从遥感图像中学习符合数据分布的模型，能够满足对高维、海量大数据的处理需求。该文用SSD作为目标识别的模型，首先根据要识别的目标物体和其对应的环境特征，利用Google Earth影像等构建多尺度、多类别的样本数据库;其次结合Tensorflow深度学习环境，在VOC2012预训练的权重文件基础上，用SSD模型对样本库进行优化训练，获取可供用于目标识别的检测模型;最后基于FLASK搭建目标识别处理页面，可根据需求输入影像，通过按钮进行目标识别处理，最终保存目标检测的后影像。利用深度学习进行遥感影像的目标识别，可主动学习目标的特征，具有较好的检测效果，同时基于FLASK将目标识别过程可视化，有助于方法的拓展应用。

关键字：遥感影像 深度学习SSD 目标识别 FLASK

Research on Remote Sensing Image Target Detection Based on Deep Learning

LI Shijia

（Stargis （Tianjin） Technology Development Co.， Ltd.， Tianjin， 300384 China）

Abstract： Remote sensing image target detection and recognition is a research difficulty at home and abroad in recent years.It uses the target recognition technology of deep learning to directly learn the model in line with the data distribution from remote sensing images， which can meet the processing needs of high-dimensional and massive big data. In this paper， SSD is used as the model of target recognition. Firstly， according to the target detection and its corresponding environmental characteristics to be identified， using Google Earth images constructs multi-scale and multi-class sample database. Then， based on the Tensorflow deep-learning environment， using SSD optimizes sample database with the weight files of VOC2012 pre-training to obtain the detection model for target recognition. Finally， the target recognition processing page is built based on FLASK， and the image can be input according to demand with button for target recognition processing to save the final image of the target detection. Using deep-learn for target recognition of remote sensing images can learn the features of the target actively and effectively. At the same time， it is useful for expanding the application with visualization of target recognition process based on FLASK.

Key Words： Remote sensing image; Deep learning; SSD;Target detection; FLASK

精准识别遥感影像中的目标物体，定位目标的空间位置，长期以来受到国内外学者的广泛关注。随着高分辨率卫星的迅猛发展，高分辨率遥感影像数据急剧增加，分辨率亦可达亚米级，从而为利用遥感数据精确分析地物目标提供了可能。通过人工解译的手段识别遥感影像中的目标，工作量大、效率低，难以满足日常工作的需求，因此研究基于大数据遥感影像的目标识别算法已成为当前的迫切需求。目前，目标检测主要分为基于特征的机器学习的传统检测算法以及基于深度学习的检测算法[1]。……