Lightweight Multiscale Feature Fusion and Multireceptive Field Feature Enhancement for Small Object Detection in the Aerial Images

Small uncrewed autonmous vehicles (UAVs) equipped with deep learning models are increasingly used to detect small objects both on the ground and in aerial environments. Since small objects occupy fewer pixels in the images, deep learning models capable of effectively extracting their features demand significant computational resources. However, the limited computational capacity of lightweight UAVs results in lower object recognition accuracy. Existing lightweight methods reduce the parameters by using lightweight convolution or an optimized feature fusion network architecture, but the former sacrifices the ability to capture the detailed features of small objects, and the latter shows a significant negative correlation between the number of parameters and detection accuracy. Aiming at the problem that it is difficult to balance the model parameters and the detection accuracy in a complex working environment, this article proposes a lightweight small object detection algorithm [lightweight multiscale feature fusion and multireceptive field feature enhancement (LMFF-MFFE)] based on multiscale feature fusion and multireceptive field feature enhancement. First, this article introduces a lightweight multiscale dense feature fusion network (MDFFN), which reduces the parameters while enriching the object feature representation by trimming and optimizing the multiscale dense feature propagation path. Second, a multireceptive field feature enhancement (MRFFE) module is integrated to capture local contextual information around objects to further enhance the effectiveness of fused multiscale features. Experimental results demonstrate superior performance on benchmark datasets: LMFF-MFFE achieves 44.5% mAP (the average precision value) on VisDrone2019 and 92.6% mAP on NWPU VHR-10, and the number of parameters is reduced by 20.4%, outperforming both baseline models and mainstream methods. The proposed LMFF-MFFE algorithm effectively balances computational efficiency and detection accuracy under resource-constrained UAV platforms, showing particular advantages in complex environments.