Small object detection algorithm based on receptive field enhancement and semantic alignment
摘要
[目的]为应对无人机高空视角下小目标检测中存在的漏检、误检及复杂背景干扰等关键技术挑战,该文提出一种基于感受野增强与语义对齐的小目标检测模型CC-YOLO.[方法]首先,引入重参数化感受野注意力卷积,构建多尺度感受野注意力卷积,通过双分支拓扑设计在浅层特征提取阶段实现大感受野建模,显著增强特征表示的空间感知能力与多场景适应性.其次,重构检测头的层级结构,优化特征传播路径,提升对极小目标的特征响应能力,同时有效压缩模型冗余参数,降低在边缘计算平台上的资源消耗.最后,提出新型特征融合策略,通过协同建模局部细节特征与全局上下文信息,缩小深层与浅层特征之间的语义差异,促进跨尺度特征在网络中的高效传递,从而提升模型对多尺度目标的检测鲁棒性.[结果]基于VisDrone 2019数据集的大量实验结果表明,CC-YOLO在检测精度与推理速度方面均优于主流实时检测模型,更适用于无人机航拍场景下的小目标检测任务.[结论]在TinyPerson与HIT-UAV数据集上开展泛化性验证实验,结果表明所提模型具备卓越的跨数据集泛化性能.
[Objective]To address the key technical challenges such as missed detection,false detection and complex background interference in small target detection from the high-altitude perspective of unmanned aerial vehicles,this paper proposes a small target detection model CC-YOLO based on receptive field enhancement and semantic alignment.[Method]Firstly,a re-parameterized receptive field attention convolution is introduced to construct a multi-scale receptive field attention convolution.Through a dual-branch topology design,large receptive field modeling is achieved in the shallow feature extraction stage,significantly enhancing the spatial perception ability and multi-scenario adaptability of feature representation.Secondly,reconstruct the hierarchical structure of the detection head,optimize the feature propagation path,enhance the feature response capability for extremely small targets,and effectively compress the redundant parameters of the model to reduce resource consumption on the edge computing platform.Finally,a new feature fusion strategy is proposed.By collaboratively modeling local detail features and global context information,the semantic differences between deep and shallow features are narrowed,promoting the efficient transmission of cross-scale features in the network,thereby enhancing the detection robustness of the model for multi-scale targets.[Result]A large number of experimental results based on the VisDrone 2019 dataset show that CC-YOLO outperforms mainstream real-time detection models in terms of detection accuracy and inference speed,and is more suitable for small target detection tasks in unmanned aerial vehicle(UAV)aerial photography scenarios.[Conclusion]Generalization verification experiments were conducted on the TinyPerson and HIT-UAV datasets.The results show that the proposed model has excellent cross-dataset generalization performance.