Research On Uav Image Object Detection Algorithm Based On Deep Learning

Using Unmanned Aerial Vehicle(UAV)for image acquisition and object detection has an important application prospect in the fields of national defense,agriculture,transportation,electric power,etc,so it has important research value.The characteristics of UAV image object detection task are large field of vision and small potential object.Classical object detection algorithms are mainly aiming at regular images,and the effect for detecting small objects in the UAV images is not acceptable.On the other hand,the deep learning based object detection algorithm has a large volume,which is difficult to meet the real-time requirements of UAV tasks.In this study,a object detection algorithm for UAV images is constructed by improving the deep learning method to improve the accuracy of detecting small objects and the real-time performance.The research mainly includes It mainly the following three aspects:(1)Analyze and preprocess the UAV visual angle image data set.The UAV image Vis Drone data set is compared with Pascal VOC and MS COCO two benchmark data sets,and the differences between the UAV data set and the conventional scene data set are comprehensively analyzed.In this thesis,the annotation information of UAV data set is visualized and the data characteristics of the object are studied in detail.In order to provide high-quality data and preparation work for the follow-up object detection,de-atomization and annotation format conversion of UAV data sets are carried out.(2)To solve the problem of poor real-time performance of the deep learning based object detection algorithm,a lightweight backbone network is constructed.The YOLOv3 object detection algorithm with a balanced performance in detection speed and accuracy is selected as the basic network.The research uses CSPNet and Bottleneck to build backbone networks,and use Ghost module to lightweight backbone networks.(3)Aiming at the problem of low accuracy of small object detection in UAV images,the anchor suitable for small object detection is redesigned.The multi-scale feature fusion strategy is optimized and high resolution feature layer is added to improve small object detection.At the same time,the attention mechanism module of channel and feature fusion are constructed to enhance the feature of small objects.The improved backbone network in this research adopts a lightweight method,which effectively reduces the total number of parameters of the model.Thus the proposed method has the potential to be transplanted to the UAV platform.The multi-scale feature fusion method designed for small objects in UAV image and the addition of high-resolution feature layer can effectively improve the model accuracy.The constructed attention mechanism module can incrementally improve the detection accuracy of small objects.Based on the improved model,the software platform of UAV image target detection is deployed.The function of rapid preview and result statistics of the image target detection results under UAV perspective is realized.