Partial feature reparameterization and shallow-level interaction for remote sensing object detection

Remote sensing object detection has recently emerged as one of the challenging topics in the field of deep learning applications due to the demand for both high detection performance and computational efficiency. To address these problems, this study introduces an efficient one-stage object detector that is designed mainly for detecting objects on remote sensing images, which consists of several innovations. Firstly, an extraction block is proposed called PRepConvBlock that leverages reparameterization convolution and partial feature utilization to effectively reduce the complexity in convolution operations, allowing for the utilization of larger kernel sizes in order to form the longer interactions between features and significantly expand receptive fields. Secondly, a unique shallow multi-scale fusion framework called SB-FPN based on Bi-FPN that utilizes the cross-interaction between shallow scale and deeper scale while inheriting the bidirectional connection from Bi-FPN to enhance the visual representation of features. Lastly, a Shallow-level Optimized Reparameterization Architecture Detector (SORA-DET) is proposed by applying several introduced innovations. This object detector is designed for UAV remote sensing object detection tasks that employ up to four detection heads. As a result, our proposed detector obtains a competitive performance that outperforms most of the other large-size models and SOTA works. In detail, the SORA-DET achieves 39.3% mAP50 in the VisDrone2019 test set while reaching up to 84.0% mAP50 in the SeaDroneSeeV2 validation set. Furthermore, our proposed detector is smaller than nearly 88.1% in parameters and has an inference speed of only 5.4 ms compared to other large-scale one-stage detectors.