Abstract
Uncrewed aerial vehicles (UAVs), leveraging their flexible low-altitude imaging and real-time data acquisition capabilities, are widely employed for aerial imagery tasks. However, UAV aerial images contain numerous small objects characterized by low pixel proportions and weak features, making them interference to background noise. To address this challenge, this paper proposes a UAV Aerial Imagery Background Suppression and Object Enhancement Detector (BSOEDet). Specifically, an Object and Background Aware Feature Extractor (OBAE) was developed, in which an auxiliary region classifier differentiates between background and object regions, and self-attention and channel attention mechanisms are combined to suppress irrelevant background and enhance object features, thereby maximizing the representation of small object feature. Furthermore, we propose a Deep Semantic Mining Small Object Module (DSMSOM). DSMSOM employs cross-attention to establish cross-spatial semantic links between deep and shallow features. By leveraging scale-invariant features between large and small objects, and semantic correlations among different objects, it mines potential small object features in the shallow layers using the semantic information of deep large objects, and further reinforces these representations with Direction Adaptive Convolution (DAConv). Moreover, a Global–Local Object Distribution Aware Classification Loss ( ${\mathcal {L}}_{\mathrm {GLOD}}$ ) is incorporated, in which the average object area and number of the entire train dataset and each batch are statistically computed to dynamically generate classification loss weights, thus enhancing the optimization contribution of small objects and few-shot classes during training. BSOEDet is validated on the VisDrone and UAVDT datasets. n the overall performance metric $\mathrm {mAP}_{50}$ , BSOEDet achieves improvements of 6.5% and 9.4% over the baseline models, and on the small object performance metric $\mathrm {mAP}_{s}$ , improvements of 4.6% and 8.4% are observed, demonstrating the effectiveness and robustness of these proposed approaches.
Funding Information
  • National Natural Science Foundation of China (U2341223)

This publication has 40 references indexed in Scilit: