Object Detection Using Scalable Feature Maps in Remote Sensing Images

In recent years, as deep learning-based general object detection has undergone continuous advancement, remote sensing object detection has become a highly concerned computer vision task. Unlike ordinary images, remote sensing images are characterized by the prevalence of small objects and complex background. Therefore, directly applying general object detection frameworks to remote sensing images often fails to achieve satisfactory detection performance. In this paper, we introduce an object detector using scalable feature maps (SCFDet) for remote sensing object detection, which comprises two integral modules: the feature refusion enhancement module (FREM) and the feature resolution rebuilding module (FRBM). The FREM utilizes different feature fusion approaches to generate enhanced features tailored to the specific requirements of classification and localization tasks. The enhanced feature map fed into the classification branch (with a size reduced to one-sixteenth of the original) leverages the larger receptive field of higher-level features to obtain relevant contextual information, assisting in classification. Meanwhile, the enhanced feature map fed into the localization branch (with the same size as the original) fuses lower-level features to acquire more fine-grained details, assisting in regression. The FRBM enlarges the low-resolution feature map in the localization branch back to its original size before obtaining the final classification results. The experiments demonstrate that compared to other remote sensing object detectors, our proposed SFMDet exhibits excellent performance on the DOTA-v1.0 dataset.