Research On Sewage Draining Exit Object Detection Of UAV Aerial Image

Object detection of Unmanned Aerial Vehicle(UAV)aerial image has a wide range of applications in many fields.This article selects the YOLOv3 algorithm to deal with the sewage draining exit object detection in UAV aerial images,which performs well in accuracy and detection speed.However,YOLOv3 model brings two problems to the sewage draining exit object detection in UAV aerial images.On one hand,compared with the general object detection datasets,the size of sewage draining exit in the UAV aerial image is relatively small,which affects the detection accurarcy in practice.To alleviate this problem,an improved Pa-YOLOv3 network is proposed,which utilizes the lower-dimensional feature map to detect small objects and increases the spatial informations of high-level feature maps through bottom-up feature fusion.On the other hand,since the payload and hardware performance of UAV platform are limited,it is necessary to make the model smaller.To this end,a channel pruning method is used to reduce the parameters,calculated amount and storage of Pa-YOLOv3.The main contributions of this article are summarized into the following folds:(1)To solve the problems of YOLOv3 algorithm in sewage draining exit object detection of UAV aerial image,including low accuracy,miss detection and false detection,an improved Pa-YOLOv3 algorithm is proposed.The size of the sewage draining exit objects in UAV aerial image is smaller than the objects in the general datasets,and the detection layers of YOLOv3 contain less spatial information,which is not conducive to the accurate detection of UAV aerial images.To alleviate this problems,this article uses 4×downsampling feature map to replace 8×downsampling feature map to improve the spatial imformation of data.Moreover,this article uses bottom-up feature fusion to enhance the spatial information of high-level feature maps.This article collects the sewage draining exit datasets for the performance evaluation of the algorithm.On the sewage draining exit datasets,the detection accuracy mAP of the improved Pa-YOLOv3 is 0.8768,which is 3.6%higher than 0.8407 using the original YOLOv3.In fact,Pa-YOLOv3 increases the model complexity to obtain the superb effectiveness,from the efficiency results of experiments,the detection speed is 16.95 FPS on GTX1080ti,which is 6.09%lower than the 18.05 FPS of original YOLOv3.In order to evaluate the generalization of Pa-YOLOv3,this article performs the algorithm on RSOD datasets,and the detection accuracy mAP of Pa-YOLOv3 is 0.9017 which is 1.6%higher than YOLOv3.(2)Due to the poor hardware performance of UAV,the lightweight model is more suitable for it.In this article,a channel pruning method was used to reduce the volume and storage of Pa-YOLOv3 elegantly.Because a large number of batch normalization layers are used in Pa-YOLOv3 model,this article uses the scaling factors in batch normalization layers as criterions to evaluate the importance of the channels.By adding regularization to the scaling factors in the batch normalization layers,the scaling factors become sparse.Then,by pruning the convolution kernels of the channels with small scaling factors,the width and size of model are both reduced.Moreover,on the sewage draining exit datasets,the number of parameters is 1.65×107 which is 78.3%lower than 7.60×107 of the original Pa-YOLOv3.And the calculated amount is 24.88 BFLOPS,which is 44.8%lower than 45.09 BFLOPS in Pa-YOLOv3.Furthermore,the required memory space of the entire model on the hard disk is 64MB,which is 78.0%lower than the 291MB of Pa-YOLOv3.The proposed pruning method makes Pa-YOLOv3 more suitable for UAV platform.