卫星视频中目标的快速检测算法研究

Fast Target Detection Algorithm in Satellite Video

随着视频卫星的不断发展,如何在卫星视频数据中准确和快速地进行目标检测逐渐成为一个研究热点.本文从两个方面改进了单阶段的目标检测网络.针对卫星图像中目标尺寸小、分辨率低的特点,利用反卷积操作丰富目标的上下文信息,同时将对应尺度的卷积特征组合成超参特征,丰富目标的细节特征;并提出图像特征多级网格化,将不同网格化的结果进行融合,提高模型的检测准确率.根据视频卫星对地凝视成像、场景移动缓慢的特点,设计出内容一致性判别网络,通过判别结果可以省略一些冗余的检测步骤,提升整体的检测效率.本实验使用"吉林一号"卫星视频数据,通过具体的实验结果分析,得出该系统对于对地凝视卫星视频中目标检测的准确率和速度都达到了较好的效果.

With the continuous development of video satellite technology, quick and accurate target detection in satellite video data has gradually become a research hotspot. This study improves the single-stage target detection framework from two aspects. In view of the features of small target size and low resolution in satellite images, the deconvolution operation is used to enrich the context information of the target, and the convolution features of the corresponding scales are combined into the super parameter features to enrich the details of the target. In addition, the image feature multilevel meshing is put forward, and the results of different meshes are fused to improve the detection accuracy of the model.According to the characteristics of satellite gaze imaging and the slow motion of the scene, we designed a content consistency discriminant network. Through the discriminant result, some redundant detection steps can be omitted to improve the overall detection efficiency. Through the concrete analysis of the experimental results of the ＂Jilin-1＂Satellite, the accuracy and speed of target detection in satellite video were achieved by the detection system.