基于改进的YOLOv10的焚烧目标检测算法

Improved Incineration Target Detection Algorithm of YOLOv10

摘要: 针对复杂焚烧场景下目标检测易受背景干扰、目标形变及光照突变影响的问题，本研究提出一种基于YOLOv10的增强检测模型。通过引入Gather-and-Distribute机制重构多尺度特征融合路径，以可变形卷积(DCNv4)提升对非刚性目标的几何表征能力，并采用Wise-Outer-MPDIoU损失函数实现基于目标几何属性的边界框优化。在自建焚烧数据集上的实验表明，所提出方法在保持高推理速度(359 FPS)的同时，平均精度(mAP)达到85.7%，较基线模型提升2.1%，显著增强了对焚烧目标的鲁棒感知与定位性能，具有良好的学术价值与工程应用潜力。

Abstract: Addressing the challenges of target detection in complex incineration scenes, which are prone to background interference, target deformation, and sudden illumination changes, this study proposes an enhanced detection model based on YOLOv10. The model incorporates a Gather-and-Distribute mechanism to reconstruct multi-scale feature fusion paths, employs deformable convolution (DCNv4) to enhance geometric representation of non-rigid targets, and adopts a Wise-Outer-MPDIoU loss function for geometry-aware bounding box optimization. Experimental results on a self-constructed incineration dataset demonstrate that the proposed method achieves a mean average precision (mAP) of 85.7% while maintaining high inference speed (359 FPS), representing a 2.1% improvement over the baseline model. The approach significantly enhances robust perception and localization of incineration targets, demonstrating considerable academic value and engineering application potential.