基于改进的YOLOV5遥感图像目标检测

摘要随着遥感卫星和航空飞行器的快速发展，遥感图像技术在道路交通检测、自然环境监控等方向中起着至关重要的作用。其中，被广泛用于对遥感图像目标的位置和类别进行识别的技术主要是遥感图像目标检测方法。但是，由于遥感图像在拍摄时的时间和角度不同等因素，使得拍摄同一自然场景图像的不同遥感图像存在一定的不同，这为目标检测工作带来了难题，并且遥感图像本身具有的背景复杂、目标小且聚集的特点，也导致了目标存在误检、漏检的情况。因此，为了提高YOLO5算法模型对遥感图像的检测精度和检测速度等性能，本文通过查阅与YOLOv:5相关文献，了解基于YOLOv5的遥感图像目标检测的发展和存在的不足，对使用YOLO5进行遥感图像目标检测提出添加小目标层、采用双注意力机制、改进双向特征网络的改进方案，具体内容如下：(1)由于遥感图像中的小目标分辨率较低，容易出现漏检、误检等情况，因此在模型中加入针对小目标的预测层，使得实验中飞机、油桶类别的召回率分别提升3.4%、0.3%,降低漏检概率。(2)由于遥感图像中存在图像噪声和复杂背景的问题，包含大量无用信息，本文通过在算法Neck层添加CoordAtt注意力机制，使模型更加注重关键特征信息，使神经网络获得更加丰富的特征信息，使得目标检测精确率、召回率、平均检测精度分别提升0.4%、1.6%、2.6%,且提升了算法的检测速率。(3)由于小目标特征信息较少，可识别度较低，结合BFPN模型对YOLOv5算法进行改进，加入了自下而上的特征融合路径，增强不同网络层之间的信息传递，使网络能够融合更多特征，减少特征信息丢失。关键词：YOLOV5遥感图像目标检测双注意力机制BPN双向特征网络Object Detection Based on Improved YOLOV5 Remote Sensing ImageZhangKaibin(College of Engineering,South China Agricultural University,Guangzhou 510642,China)Abstract:With the rapid development of remote sensing satellites and aviation aircraft,remotesensing image technology plays an important role in many fields such as road traffic detectionand natural environment monitoring.Among them,remote sensing image object detection iswidely used to recognize the position and category of targets in remote sensing images.However,due to factors such as different shooting times and angles of remote sensing images,there are certain differences in different remote sensing images of the same natural scene,which poses challenges for object detection.Moreover,remote sensing images themselveshave the characteristics of complex backgrounds,small targets,and clustering,which also leadto the occurrence of false or missed detection of targets.Therefore,in order to improve thedetection accuracy and speed of YOLOv5 algorithm model for remote sensing images,thisarticle reviews relevant literature on YOLOv5 to understand the development andshortcomings of YOLOv5 based remote sensing image object detection.For remote sensingimage object detection using YOLOv5,an improvement plan is proposed,which includesadding a small target layer,using a dual attention mechanism,and improving a bidirectionalfeature network.The specific content is as follows:(1)Due to the low resolution of small targets in remote sensing images,they are prone tomissed and false detections.Therefore,adding a prediction layer for small targets to the modelincreases the recall rates of aircraft and oil tank categories by 3.4%and 0.3%respectively inthe experiment,reducing the probability of missed detections.(2)Due to the problems of image noise and complex background in remote sensing images,which contain a large amount of useless information,this article adds CoordAtt attentionmechanism to the algorithm Neck layer to make the model pay more attention to key featureinformation,enabling the neural network to obtain richer feature information.This improvesthe accuracy,recall,and average detection accuracy of target detection by 0.4%,1.6%,and2.6%,respectively,and improves the detection rate of the algorithm.(3)Due to the