In order to solve the problems of poor adaptability, low efficiency and accurate prior knowledge in matching image pair extraction in 3D reconstruction of UAV images, this paper proposes a matching image pair extraction method based on feature expression.Firstly, SiftGPU algorithm is used to extract SIFT features from a group of UAV images. Secondly, a hierarchical vocabulary tree is constructed for SIFT feature vector set, and the comprehensive weight factors of each image and each image in the image set are calculated and sorted according to the score size. Finally, the image retrieval threshold is calculated adaptively to obtain the final image matching image pair. Through the experiments of UAV images with different topography and landforms, this method can quickly and effectively obtain the matching image pairs of the query images. Compared with the conventional vocabulary tree retrieval method, the precision rate is increased by 19% ~ 24%, and the recall rate is increased by 17.4% ~ 30.9%.Especially for massive UAV images data, this method has higher processing efficiency.