Unmanned aerial vehicle (UAV) aerial image object detection is a valuable and challenging research field. Despite the breakthrough of deep learning-based object detection networks in natural scenes, UAV images often exhibit characteristics such as a high proportion of small objects, dense distribution, and significant variations in object scales, posing great challenges for accurate detection. To address these issues, we propose an innovative multiscale feature fusion small object detection network (MFFSODNet). First, concerning the high proportion of small objects in UAV images, an additional tiny object prediction head is introduced instead of the large object prediction head. This approach provides a good detection accuracy of small objects and significantly reduces the parameters. Second, to enhance the feature extraction capability of the network for fine-grained information from small objects, a multiscale feature extraction module (MSFEM) is designed, which could extract rich and valuable multiscale feature information through convolution operation of different scales on multiple branches. Third, to fuse the fine-grained information from shallow feature maps and the semantic information from deep feature maps, a new bidirectional dense feature pyramid network (BDFPN) is proposed. By expanding the feature pyramid network scale and introducing skip connections, BDFPN achieves efficient multiscale information fusion. Extensive experiments on the VisDrone and UAVDT benchmark datasets demonstrate that MFFSODNet outperforms the state-of-the-art object detection methods and further validate the effectiveness and generalization of MFFSODNet on photovoltaic array defect datasets (PVDs).

中文翻译：

---

MFFSODNet：无人机航拍图像的多尺度特征融合小物体检测网络

无人机（UAV）航空图像目标检测是一个有价值且具有挑战性的研究领域。尽管基于深度学习的目标检测网络在自然场景中取得了突破，但无人机图像往往呈现出小目标比例高、分布密集、目标尺度变化大等特点，给准确检测带来了巨大挑战。为了解决这些问题，我们提出了一种创新的多尺度特征融合小物体检测网络（MFFSODNet）。首先，针对无人机图像中小物体比例较高的问题，引入了额外的小物体预测头来代替大物体预测头。该方法提供了良好的小物体检测精度并显着减少了参数。其次，为了增强网络对小物体细粒度信息的特征提取能力，设计了多尺度特征提取模块（MSFEM），通过多个分支上不同尺度的卷积运算，可以提取丰富且有价值的多尺度特征信息。第三，为了融合浅层特征图的细粒度信息和深层特征图的语义信息，提出了一种新的双向密集特征金字塔网络（BDFPN）。通过扩大特征金字塔网络规模并引入跳跃连接，BDFPN实现了高效的多尺度信息融合。在 VisDrone 和 UAVDT 基准数据集上进行的大量实验表明，MFFSODNet 优于最先进的目标检测方法，并进一步验证了 MFFSODNet 在光伏阵列缺陷数据集（PVD）上的有效性和泛化性。