基于重参数化注意力的无人机航拍目标检测方法

Detection of Object in UAV Aerial Photography Based on Reparameterized Attention

针对无人机航拍图像目标尺度变化大、目标与背景相互干扰所导致的误检、漏检问题,提出一种基于重参数化注意力的目标检测方法,并将其应用于无人机航拍目标检测。首先,通过提出重参数化坐标注意力模块来增强相关特征,提升网络对上下文信息的捕捉能力;其次,设计多尺度感受野增强模块来重构骨干网络,从而增强特征图的接受域,提升网络的特征提取能力;接着,提出四尺度特征融合检测网络,提升网络对小目标的检测能力;最后,引入解耦检测头来解决分类与回归任务之间的冲突。在VisDrone2021数据集上进行实验,所提方法的mAP0.5相比原算法提高了7.6个百分点,召回率提升了5.5个百分点,与其他方法相比也具有明显优势。实验结果证明,改进方法能够较好地解决上述误检、漏检问题,具有良好的检测效果。

Aiming at the problems of false detection and missed detection of objects in UAV aerial images caused by large changes in the object scale or mutual interference between the object and background,a object detection method based on reparameterized attention is proposed,which is applied to UAV aerial object detection.Firstly,the reparameterized coordinate attention module is proposed to enhance the relevant features and improve network ability to capture context information.Secondly,a multi-scale receptive field enhancement module is designed to reconstruct the backbone network,thereby enhancing the acceptance domain of feature map and improving the feature extraction ability of network.Then,a four-scale feature fusion detection network is proposed to improve detection ability of the network for small objects.Finally,a decoupling detection head is introduced to resolve the conflict between classification and regression tasks.In experiments on the VisDrone2021 dataset,mAP0.5 and racall rate of our algorithm is improved respectively by 7.6 and 5.5 percentage points in comparison with the original algorithm,and the algorithm also shows obvious advantages over other methods.The experimental result shows that,the improved method can better solve the above-mentioned problems of false detection and missed detection,and has good detection effect.