Small Target Detection Algorithm Of UAV Based On Deep Learning

In recent years,drones have been widely used in power inspection.Due to the complexity of the power field environment,drone operators need to rely on the drone flight technology assessment system for strict assessment before taking up their jobs.At the test site,the detection effect of the vision system on long-distance small UAVs will determine the ability of the assessment system.Due to the small size of small objects in the image and the inconspicuous object features,traditional object detection algorithms cannot effectively detect small objects.The continuous improvement of deep learning theory has continuously improved the ability of small target detection.In order to improve the detection effect of UAV small targets in the high-voltage tower scenario,the main contents of this topic include:（1）A dataset of specific UAV small targets is constructed.Since there is no publicly available UAV dataset,this paper firstly completes the image acquisition of three different types of UAVs in different scenarios,and then uses background expansion,rotation distortion and other methods to enhance data set.（2）Research the YOLO series target detection algorithm based on deep learning.In order to achieve rapid detection of UAV small targets,YOLOv3 is selected as the network architecture after comparative analysis of the structural characteristics and advantages and disadvantages of YOLO series algorithms.Then,Improve the Yolov3 algorithm from 5 aspects.79.24% is the value of detection accuracy of the algorithm in the high-voltage tower scenario increase,23.38% and 19.08% reduction in missed detection rate and false detection rate,respectively.（3）Research the RCNN series target detection algorithm based on deep learning.In order to achieve high-precision detection of UAV small targets,after comparing and analyzing four RCNN algorithms,Mask RCNN is selected as the network architecture,and the Mask RCNN algorithm is improved according to the experimental results.The improved algorithm is used in high-voltage towers.84.63% is the value of detection accuracy of the algorithm in the high-voltage tower scenario increase,13.34% and16.49% reduction in missed detection rate and false detection rate,respectively.（4）Design and implement the small target detection algorithm of YOLOv3+Mask RCNN.In order to further improve the detection effect of UAV small targets,the improved YOLOv3 algorithm and the improved Mask RCNN algorithm were integrated to form the YOLOv3+Mask RCNN algorithm,and according to the experimental results,the algorithm was improved again.88.76% is the value of detection accuracy of the algorithm in the high-voltage tower scenario increase,7.67% and 8.21% reduction in missed detection rate and false detection rate,respectively.