基于注意力机制的航拍图像地面目标旋转检测实验设计  

Experimental design of ground target rotation detection in aerial images based on attention mechanism

作　　者：陈远林 贺云涛 李琳琳 CHEN Yuanlin;HE Yuntao;LI Linlin(School of Astronautics,Beijing Institute of Technology,Beijing 100081,China;School of Automation,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京理工大学航天学院,北京100081 [2]北京科技大学自动化学院,北京100083

出　　处：《实验技术与管理》2024年第6期57-63,共7页Experimental Technology and Management

摘　　要：针对无人机航拍图像中目标实例存在遮挡、旋转、尺度变化等情况,该文设计了一种基于注意力机制的旋转目标检测实验设计方案。首先,对开源DOTAv-1.0数据集样本图像进行了图像增强和裁剪,构造二次数据集,使整体实验样本数量提升11.72倍。然后,通过在S2A-Net网络的特征融合网络部分嵌入CBAM注意力机制模块,增强对输入特征空间和通道的有效信息利用率。最后,通过消融实验验证了CBAM模块的有效性,整体性能提升1.3%,对航拍小目标实例small-vehicle的检测精度mAP值由0.402 9提升到0.539 8,提升了34%,检测速率达到24 fps。[Objects]Object detection algorithms have shown promising results in drone aerial photography detection and dynamic tracking.However,challenges arise during drone aerial photography owing to occlusion,rotation,and scale changes in captured target instances.The traditional YOLO object detection algorithm is not ideal for detecting rotating targets effectively.[Methods]To address this issue,this article introduces an experimental ground target rotation detection algorithm.It incorporates the CBAM attention mechanism into the S2ANet algorithm,offering improved solutions to mixed background occlusion,target instance rotation and tilt,and multi-scale image changes in aerial imagery.Prior to experimentation,preprocessing is conducted to expand the DOTA dataset.The experimental dataset’s sample images have been increased from 2806 to 4209 through image rotation and translation techniques.To facilitate the training and processing of deep learning algorithms,all images are standardized to a resolution of 1024×1024.Further cropping and enhancing of the expanded samples result in the creation of the DOTA_split dataset,containing a total of 32,877 samples.Following experimental preprocessing,the overall sample size increases to 11.72 times the original,significantly enhancing the accuracy of neural network training.Subsequently,the S2ANet network is chosen as the rotation object detection model,with RetinaNet serving as the baseline network model.The backbone network,ResNet,is paired with the neck network,FPN,while the head detection network consists of two FCN subnetworks.Additionally,the S2ANet network model incorporates the feature alignment network and the orientation detection module.Furthermore,a CBAM attention mechanism module is integrated into the feature fusion section of the S2ANet network model to enhance the effective utilization of input feature space and channels.Through the embedding of a CBAM attention mechanism network into each layer of the fused feature map obtained via upsampling in the FPN layer,the

关 键 词：深度学习 航空遥感 倾斜检测 注意力机制 实验设计 

分 类 号：V279[航空宇航科学与技术—飞行器设计]