Research And Application Of Improved YOLOv3 Algorithm

The development of remote sensing technology drives the application of remote sensing images more and more widely,and remote sensing image target detection is widely used in military combat,medical health,traffic safety and agriculture and other military and civilian production and life fields.Target detection research aims to improve the accuracy and real-time performance of target detection.Traditional target detection algorithms mostly extract features by manual design,which affects the performance and application effect of target detection.In recent years,with the deepening of machine learning and data mining research,there are several deep learning-based target detection algorithms.The purpose and core research of this paper is to improve the accuracy of target detection based on deep learning target detection algorithm YOLOv3 in detecting target densities,optimizing for the existence of small target detection such as missed detection,false detection and other poor detection results.Firstly,data enhancement is used to expand the dataset in order to address the problem of missed detection and false detection that can occur in single-stage models like YOLO,to create a semantically rich dataset,and to screen out invalid data in the dataset based on the aspect ratio in order to improve the speed and accuracy of model training in the network.Secondly,for the problem that YOLOv3’s small target detection effect is lacking,the original three feature scales of YOLOv3 are increased to four,thus reducing the missed detection rate of small target objects,and for the problem that the accuracy of YOLOv3’s bounding boxes needs to be improved,new candidate boxes for target detection of agricultural machinery are obtained by using improved K-Means target box clustering,thus improving the For the problem that the image samples are similar in size and the value of the perceptual field decreases after multiple convolutions,the problem is solved by fusing the improved scale scaling method to retain the multi-scale advantage.Finally,based on the improved scale scaling,the backbone network is optimized by adding a hybrid attention module and redistributing the number of residual blocks,and a new network model,namely "YOLO-KMR-Attention",is proposed.By comparing with the classical target detection network model,the m AP of the proposed model is improved from 91.1% to 92.1%,and the detection time of a single image is only 21.8ms,which is the same as the detection time of several classical models and meets the requirement of real-time performance.In summary,the accuracy and real-time performance of the target detection algorithm model "YOLO-KMR-Attention" proposed in this paper have been quantified and improved to a certain extent,which shows that the innovative improvement of this paper is effective and has certain research value,and the use of the "YOLO-KMR-Attention" has been improved to a certain extent.KMR-Attention" algorithm model is designed and implemented to detect the target of agricultural machinery,which has some application value for agricultural production.