基于模型压缩的ED-YOLO电力巡检无人机避障目标检测算法

ED-YOLO power inspection UAV obstacle avoidance target detection algorithm based on model compression

针对现有卷积神经网络模型体积大、运算量高,导致电力巡检无人机检测速率与精度无法兼顾的问题,提出一种基于模型压缩的ED-YOLO网络实现无人机避障的目标检测算法。该目标检测算法以YOLOv4为基础,首先在主干网络中加入通道注意力机制,在不增加计算量前提下提高检测精度;其次在特征金字塔部分运用深度可分离卷积替换传统卷积,减少卷积计算量;最后利用模型压缩策略裁剪网络中冗余通道,减小模型体积并提高模型检测速度。在自主构建的9600张电力巡检无人机飞行障碍的数据集进行测试,ED-YOLO与YOLOv4相比,其障碍物目标检测的平均精度均值只降低了1.4%,而模型体积减少了94.9%,浮点运算量减少了82.1%,预测速度提升了2.3倍。实验结果表明,对比多种其他现存方法,本文提出的基于模型压缩的ED-YOLO目标检测算法有着精度高、体积小和检测速度快的优势,满足电力巡检无机避障检测要求。

Aiming at the problem that existing convolutional neural network models are large in size and high in computation,which results in not being able to consider both detection rate and accuracy of power inspection UAVs,an ED-YOLO network based on model compression is proposed to achieve the target detection algorithm for UAV obstacle avoidance.The target detection algorithm is based on YOLOv4,which firstly adds a channel attention mechanism to the backbone network to improve detection accuracy without increasing the amount of computation.Secondly,the depth separable convolution is used to replace the traditional convolution in the feature pyramid part to reduce the amount of convolutional computation.Finally,the model compression strategy is used to trim the redundant channels in the network to reduce the model size and improve the model detection speed.Tests were conducted on the dataset independently constructed with 9600 flight obstacles of power inspection UAV,the obstacle target average detection accuracy for ED-YOLO is reduced only by 1.4%compared with that for YOLOv4,while the model size is reduced by 94.9%,the amount of floating point operations is reduced by 82.1%and the prediction speed is increased by 2.3 times.Experiment results show that compared with various other existing methods,the ED-YOLO target detection algorithm based on model compression proposed in this paper has the advantages of high accuracy,small size and fast detection speed,and meets the requirements of obstacle avoidance detection for power inspection UAVs.