Design And Implementation Of A Target Detection Algorithm For Aerial Images Based On Deep Learning

The object detection and recognition of aerial images is a very important task,which is widely used in the fields of map mapping,disaster prediction and management,agricultural detection,urban planning and construction.With the rapid development of deep learning technology,the aerial image object detection algorithm based on deep neural network has achieved far more effects than traditional object detection algorithms.However,when these deep learning algorithms are applied to actual aerial photography detection Systems,they often face the problem of excessive computational overhead.To achieve real-time detection requires the use of multiple high-performance GPUs,and the huge computing power requirements limit the application of the algorithm in real-time detection tasks of aerial images.Therefore,in order to solve the problem of large amount of calculation and poor real-time performance of the object detection algorithm for aerial images,this article has made two improvements at the algorithm and hardware level:1.At the algorithm level,the YOLOv4 algorithm,which has a very balanced speed and accuracy so far,is selected as the benchmark algorithm,and it is improved and optimized for specific aerial image object detection tasks.We first use the K-means++ clustering algorithm to redesign the initial value of the anchor frame to improve the detection accuracy of the aerial image object detection task.At the same time,we have also improved the SPP module and activation function in the YOLOv4 network to make it more suitable for deployment on the FPGA platform.These improved measures further reduce the amount of calculations and achieve faster operating speeds when the accuracy is very small.2.At the hardware implementation level,Xilinx FPGA is selected as the hardware implementation device.Use Xilinx's Vitis AI development platform and DPU IP Core to build a dedicated neural network accelerator.With a series of open source tool chains,the model was quantified,compiled to generate DPU-specific instructions,and deployed to the ZCU102 platform.The Experimental results show that after the algorithm and implementation technology designed in this paper are deployed on the ZCU102 platform,the main module and DSP clock can reach operating speeds of 350 MHz and 650 MHz,respectively,it can achieve about 2.7T of calculation per second.For the input image of 416×416,it can process about 33 frames per second,and the power consumption is about 22 W.Compared with previous research results,it has a better energy efficiency ratio.In terms of target detection accuracy,m AP based on the DOTA data set reaches 58.7%,which can better meet actual application requirements.