Research On Multiscale Detection Method Of UAV Target In Optical Image

At present,the technology of UAV is becoming more and more mature,and the application range of UAV is also increasing,and some hidden dangers also follow.Some people's illegal use of drones has resulted in safety accidents.Therefore,it is necessary to control UAV flight reasonably.Infrared camera and visible light camera are common UAV monitoring means.In this paper,the UAV target in infrared and visible light image is detected in multi-scale to increase the monitoring range of the detector to UAV.Firstly,the target scale boundary analysis of the target detection algorithm yolov3 is studied.The basic structure of the first-order target detection algorithm based on anchor frame and the basic principle of yolov3 are introduced.Bilinear interpolation instead of nearest neighbor interpolation is used to enhance the detection ability of yolov3 for small targets.After training with UAV image set,the target scale boundary of yolov3 is tested.It is found that bilinear interpolation can not improve the detection ability of yolov3 for small targets.Secondly,based on darknet-53,the main network of yolov3,the detection method of small target UAV in image is studied,which provides theoretical and experimental basis for improving the realization of multi-scale detection of UAV target in yolov3.The possibility of using bounding box regression to detect small targets is studied.It is concluded that when using bounding box to detect small targets,not only there are many parameters,but also the performance of small target detection must be suppressed because the contribution of large-scale targets to loss function is greater than that of small targets.To solve the problem that bounding box regression is not suitable for small target detection,a network based on scale-transfer layer is designed to predict the probability that every point in the image is a small target.In order to verify the effect of receptive field on small target detection performance,different depth parts of darknet-53 are intercepted as feature extraction network.According to the characteristics of small targets,a data enhancement method is designed.Regularization and batch normalization are used to suppress over fitting in the network.Using the collected and simulated small target images,the designed small target detection method is tested.Finally,in order to reduce the minimum detectable scale of Yolov3,a multi-scale target detection algorithm based on feature fusion is designed.In order to detect multi-scale targets by boundary box regression and case segmentation,the feature extraction network of Yolov3 is added with shallow branches to detect small targets,and the multi task loss function is designed.In convolutional neural network,the shallow features have more location information and the deep features have more semantic information.When making data sets,different annotation methods are used for different scale targets,and the targets smaller than the set scale are marked with pixel sets.Using image to train and test the designed multi-scale target detection method,the detection rate is 97.78% when the false alarm rate of small target image sequence is 1.6 × 10^-4,and the AP of UAV is increased by 1.8% compared with Yolov3.