An Efficient and Robust Integrated Geospatial Object Detection Framework for High Spatial Resolution Remote Sensing Imagery

摘要: Geospatial object detection from high spatial resolution (HSR) remote sensing imagery is a significant and challenging problem when further analyzing object-related information for civil engineering applications. However, the computational efficiency separate region generation localization steps are two big obstacles performance improvement of traditional convolutional neural network (CNN)-based methods. Although recent methods based on CNN can extract features automatically, these still feature extraction stages, resulting in time consumption low efficiency. As influencing factor, acquisition large quantity manually annotated samples HSR objects requires expert experience, which expensive unreliable. Despite progress made natural image fields, complex distribution makes it difficult to directly deal with task. To solve above problems, highly efficient robust integrated geospatial framework faster region-based (Faster R-CNN) proposed this paper. The method realizes procedure by sharing between proposal stage stage. In addition, pre-training mechanism utilized improve multi-class transfer learning domain domain. Extensive experiments comprehensive evaluations publicly available 10-class dataset were conducted evaluate method.