基于改进YOLOv5的机场目标检测方法

An improved YOLOv5-based target detection method for airports

针对机场复杂场景监视能力差的问题,提出了一种基于改进YOLOv5的机场场面目标检测方法。首先,在模型上以跨特征的CSPDarknet作为主干特征提取网络,能大幅减少模型参数量,同时,加入Focal loss以缓解正负样本不平衡的状况,加入ECA注意力机制以加强特征间联系。其次,在数据集上采用CycleGAN进行夜间场景的风格迁移,以提升复杂背景下的检测能力。再次,在训练策略上采用mosaic数据增强,标签平滑正则,余弦退火学习率下降,并以Adam作为优化器加速模型收敛。最后,结合改进后的航空遥感数据集进行实例验证,结果表明,所提算法的MAP值较原YOLOv5s提高了2.79%,其中精确率提高了2.55%,召回率提高了0.49%,检测速度和模型参数量变化不大。

Aiming at the problem of poor monitoring ability in airport complex scenes,an airport scene target detection method based on improved YOLOv5 was proposed.Firstly,the cross-feature CSPDarknet is used as the backbone feature extraction network in the model,which can significantly reduce the number of model parameters.At the same time,Focal loss is added to alleviate the imbalance of positive and negative samples,and the ECA attention mechanism is added to strengthen the inter-feature connection.Secondly,CycleGAN is used in the dataset for the style migration of nighttime scenes to improve the detection ability in complex backgrounds.After that,mosaic data enhancement,label smoothing regularity,cosine annealing learning rate reduction,and Adam as an optimizer are used in the training strategy to accelerate model convergence.Finally,an example validation is performed with the improved aerial remote sensing dataset,and the results show that the MAP value of the proposed algorithm is improved by 2.79%compared with the original YOLOv5s,in which the accuracy rate is improved by 2.55%and the recall rate is improved by 0.49%,with little change in the detection speed and the number of model parameters,which facilitates the improvement of airport surveillance capabilities.