An Efficient Aerial Image Detection with Variable Receptive Fields
Wenbin Liu, Liangren Shi, Guocheng An
Abstract
This article presents VRF-DETR, a lightweight real-time object detection framework for aerial remote sensing images, aimed at addressing the challenge of insufficient receptive fields for easily confused categories due to differences in height and perspective. Based on the RT-DETR architecture, our approach introduces three key innovations: the multi-scale receptive field adaptive fusion (MSRF2) module replaces the Transformer encoder with parallel dilated convolutions and spatial-channel attention to adjust receptive fields for confusing objects dynamically; the gated multi-scale context (GMSC) block reconstructs the backbone using Gated Multi-Scale Context units with attention-gated convolution (AGConv), reducing parameters while enhancing multi-scale feature extraction; and the context-guided fusion (CGF) module optimizes feature fusion via context-guided weighting to resolve multi-scale semantic conflicts. Evaluations were conducted on both the VisDrone2019 and UAVDT datasets, where VRF-DETR achieved the mAP50 of 52.1% and the mAP50-95 of 32.2% on the VisDrone2019 validation set, surpassing RT-DETR by 4.9% and 3.5%, respectively, while reducing parameters by 32% and FLOPs by 22%. It maintains real-time performance (62.1 FPS) and generalizes effectively, outperforming state-of-the-art methods in accuracy-efficiency trade-offs for aerial object detection.