基于NATCA-GreaterYOLO的航拍小目标检测

Small Object Detection in VisDrone ImagesBased on NATCA-Greater YOLO

为解决通用目标检测模型在无人机航拍场景下存在的物体尺度变化剧烈及复杂的背景干扰等问题,本文主要对基于NATCA-Greater YOLO的航拍小目标检测进行研究。在特征提取网络的最后一层,加入邻域注意力转换器(neighborhood attention transformer,NAT),以保留足够的全局上下文信息,并提取更多不同的特征。同时,在特征融合网络(Neck)部分,加入坐标注意力(coordinate attention,CA)模块,以获取通道信息和更长范围的位置信息,将原卷积块中的激活函数替换为Meta-ACON,并使用NAT作为新网络的预测层,以VisDrone2019-DET目标检测数据集为基准,在VisDrone2019-DET-test-dev数据集上进行测试。为了评估NATCA-Greater YOLO模型在航拍小目标检测任务中的有效性,采用Faster R-CNN、RetinaNet和单步多框目标检测(single shot multiBox detector,SSD)等检测网络在测试集上进行对比检测。研究结果表明,NATCA-Greater YOLO检测的平均精度为42%,与最先进的检测网络TPH-YOLOv5相比,NATCA-Greater YOLO的检测精度提升了2.9%,说明该模型可以准确地定位并识别目标。该研究具有一定的创新性。

In order to solve the problems of the general object detection model in the images captured by drones,including drastic scale variance and complex background interference,this paper focuses on the detection of small objects in aerial photography based on NATCA-Greater YOLO.We add neighborhood attention transformer(NAT)to the last layer of the feature extraction network to retain sufficient global context information and extract more different features.Meanwhile,in the feature fusion network(Neck)part,the coordinate attention(CA)module is added to obtain channel information and longer range location information,the activation function in the original convolutional block is replaced with Meta-ACON,and NAT is used as the prediction layer of the new network.Using the VisDrone2019-DET object detection dataset as a benchmark,tests were conducted on the VisDrone2019-DET-test-dev dataset.To evaluate the effectiveness of the NATCA-Greater YOLO model in the aerial photography small object detection task,detection networks such as Faster R-CNN,RetinaNet and SSD(single shot multiBox detector)were used for comparative testing on the test set.The results show that the average accuracy of NATCA-Greater YOLO detection is 42%,which is 2.9%improvement compared to the state-of-the-art detection network TPH-YOLOv5.This study is innovative.