基于深度学习的遥感图像微小目标检测方法研究

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

摘要:

遥感图像中含有大量的微小目标,只有准确检测到这些微小目标,才能实现远程目标的识别与跟踪。为了给远程跟踪工作提供有效的辅助工具,以深度学习算法为技术支持,优化设计遥感图像微小目标检测方法。利用硬件设备实时采集包含微小目标的遥感图像,通过几何校正、灰度化转换、噪声抑制、去雾以及图像增强等步骤,完成初始图像的预处理。通过前景与背景图像的分割,选择遥感图像中的待检测目标。构建深度卷积神经网络作为深度学习算法的运行环境,经过前向传播、反向传播提取遥感图像特征。最终通过特征匹配,得出包含微小目标数量以及位置坐标的检测结果。通过性能测试实验得出结论：与传统遥感图像目标检测方法相比,优化设计方法的查准率和查全率分别提高了6.3%和10.74%,目标位置检测误差得到明显降低,且响应时间缩短了2440ms,由此证明优化设计方法具有良好的检测性能。

Abstract:

There are a large number of small targets in remote sensing images. Accurate detection of them is the basis of remote target recognition and tracking. In order to provide effective auxiliary tools for remote tracking, the micro target detection method of remote sensing image is optimized with the technical support of deep learning algorithm. The hardware equipment is used to collect the remote sensing image containing micro targets in real time, and the preprocessing of the initial image is completed through the steps of geometric correction, gray conversion, noise suppression, defogging and image enhancement. Through the segmentation of foreground and background image, the target to be detected in remote sensing image is selected. The deep convolution neural network is constructed as the operation environment of the deep learning algorithm, and the remote sensing image features are extracted through forward propagation and back propagation. Finally, through feature matching, the detection results including the number of small targets and position coordinates are obtained. Through the performance test experiment, it is concluded that compared with the traditional remote sensing image target detection method, the precision and recall of the optimal design method are increased by 6.3% and 10.74% respectively, the target position detection error is significantly reduced, and the response time is shortened by 2440ms, which proves that the optimal design method has good detection performance.