Abstract: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.