融合位置注意力的无人机影像多目标检测方法

A multi-target detection method for UAV images fused with location attention

无人机能够在城市规划、侦察、监视等场景下,通过目标检测技术提供准确的目标位置和类别信息,为后台处理提供详细的信息,但现有方法在无人机影像检测时存在场景泛化能力不足、小目标漏检率高等问题。鉴于此,提出一种基于回归的检测方法,在骨干网络中使用位置注意力机制为正负样本特征赋权,提高模型对正样本的学习能力;构建四个输出尺度的特征图融合金字塔,并采用改进的非极大值抑制算法精准筛选最终的输出检测框。为降低正负样本不均衡带来的影像,一方面采用交叉熵损失函数,另一方面对训练数据集进行样本增强处理。实验结果表明,所提出模型在测试数据集上的检测精度明显优于对比模型,并且在不同场景下表现出良好的泛化能力,其测试速度可达到实时检测的水平。

UAV can provide accurate target location and category information through target detection technology in urban planning,reconnaissance,surveillance and other scenes,and provide detailed information for background processing.However,the existing methods have the problems of insufficient scene generalization ability and high miss rate of small targets in detecting UAV images.To solve these problems,a detection method based on regression is proposed.In the backbone network,the location attention mechanism is used to weight the positive and negative sample features to improve the learning ability of the model to the positive samples;A feature map fusion pyramid with four output scales is constructed,and an improved non maximum suppression algorithm is used to precisely screen the final output detection box.In order to reduce the image caused by the imbalance of positive and negative samples,on the one hand,the cross entropy loss function is used,and on the other hand,the training data set is enhanced.The experimental results show that the proposed model has significantly better detection accuracy than the comparison model on the test data set,and has good generalization ability in different scenarios,and its test speed has reached the level of real-time detection.