DRE-YOLO

Detection of Unexploded Ordnance (UXO) based on drone platforms is crucial for public safety; however, it faces three significant challenges in complex outdoor environments: abrupt variations in target scale, degradation of edge information due to vegetation occlusion, and a high false detection rate caused by the similarity in texture between targets and backgrounds.Existing methods for detecting rotating targets exhibit inherent limitations: the traditional fixed receptive field design results in insufficient discriminability of small target features; in scenarios with vegetation occlusion, upsampling operations struggle to mitigate the degradation of geometric information at target edges; and significant background noise substantially reduces detection reliability.To address these challenges, this paper introduces DRE-YOLO, a multi-scale dynamic refinement network framework. First, a Deformable Adaptive Module (DAM-C3) is designed to dynamically adjust the receptive field size through deformable convolutional kernels while integrating local-global contexts, significantly enhancing scale adaptability. Secondly, an Efficient Shift-Upsampling (ES-Upsample) is developed, which combines a decoupled upsampling mechanism with a channel displacement fusion layer to effectively suppress edge degradation caused by vegetation occlusion. Finally, a Dual-Polarity Attention (DPA) is introduced in the detection head, utilizing a dynamic feature scaler and differential polarity modeling to substantially mitigate interference from background textures that are similar to the targets.基于无人机平台的未爆弹药（Unexploded Ordnance, UXO）检测对于公共安全至关重要，但在复杂户外环境中面临三大显著挑战：目标尺度突变、植被遮挡导致的边缘信息退化，以及目标与背景纹理相似性引发的高误检率。现有旋转目标检测方法存在固有局限：传统固定感受野设计会导致小目标特征判别能力不足；在植被遮挡场景中，上采样操作难以缓解目标边缘的几何信息退化问题；显著的背景噪声则会大幅降低检测可靠性。为解决上述挑战，本文提出DRE-YOLO——一种多尺度动态细化网络框架。首先，设计可变形自适应模块（Deformable Adaptive Module, DAM-C3），通过可变形卷积核动态调整感受野尺寸并融合局部-全局上下文，显著提升尺度适应性；其次，提出高效移位上采样（Efficient Shift-Upsampling, ES-Upsample），将解耦上采样机制与通道移位融合层相结合，可有效抑制植被遮挡引发的边缘退化；最后，在检测头中引入双极性注意力（Dual-Polarity Attention, DPA），通过动态特征缩放器与差分极性建模，大幅削弱与目标纹理相似的背景纹理带来的干扰。