面向边缘计算平台及遥感影像的实时检测算法

针对现有目标检测算法难以满足无人机遥感中实时检测的问题，提出了一种基于ShuffleNetv2及结构化剪枝的模型压缩方法。以YOLOv5m为基础，将ShuffleNetv2模型作为YOLOv5m的主干网络，减少模型的参数量及计算量，提升模型推理速度；其次，利用ECA注意力机制替换ShuffleNetv2中的SE模块，强化主干网络的特征提取能力；再者，以FocalEIoU作为YOLOv5算法的损失函数，提升模型的回归能力；最后，利用通道剪枝算法剔除Neck结构中冗余的参数，进一步压缩模型的参数及计算量，并通过模型微调的方式提升剪枝模型的精度。实验结果表明，在相同的测试环境下，与YOLOv5m相比，本文所提出模型的参数量及浮点运算量分别降低了86.3%和80.0%,mAP@0.5和mAP@0.5:0.95达到了92%及50.4%,优于所对比的其他主流检测算法。此外，所提出的模型在AGX边缘计算平台上达到了35帧/s的检测速度，满足实时检测的要求。

Real-time detection algorithm for edge computing platforms and remote sensing imagery

To address the issue of existing object detection algorithms struggling to meet real-time detection requirements in UAV remote sensing, we propose a model compression method based on ShuffleNetv2 and structured pruning. Using YOLOv5m as the foundation, we incorporate the ShuffleNetv2 model as the backbone network of YOLOv5m, reducing the model′s parameter count and computational complexity while improving inference speed. Furthermore, we employ the ECA attention mechanism to replace the SE module in ShuffleNetv2, enhancing the feature extraction capability of the backbone network. Additionally, we adopt FocalEIoU as the loss function for the YOLOv5 algorithm, improving the model′s regression ability. Finally, we use channel pruning to eliminate redundant parameters in the Neck structure, further compressing the model′s parameters and computational complexity, and enhancing the pruned model′s accuracy through fine-tuning.Experimental results show that, under the same testing environment, compared to YOLOv5m, the proposed model reduces the parameter count and floating-point operations by 86.3% and 80.0%, respectively. The model achieves an mAP@0.5 of 92% and an mAP@0.5:0.95 of 50.4%, outperforming other mainstream detection algorithms. Moreover, the proposed model achieves a detection speed of 35 frames/s on the AGX edge computing platform, satisfying the requirements for real-time detection.