| Remote sensing is the basic method of detecting and classifying objects on earth,using the propagation and reception of electromagnetic waves(including light waves)emitted from satellites or aircraft to the ground.In recent years,with the rapid development of remote sensing technology,and the constantly improving temporal and spatial resolution of earth observation,remote sensing data present a typical big data problem,which is undoubtedly a tough challenge for the current high-performance computing(HPC)system and the corresponding solutions.Therefore,it is urgent to study novel solutions that can effectively process remote sensing data.In nearly a decade,on the other hand,deep learning at an extremely rapid development speed,and has become the most promising methods during big data era,especially represented by convolution neural network,which brings great promote research on remote sensing image analysis on the problem such as target identification and extraction of remote sensing.However,due to the much more complexity of neural network architecture,deep learning needs enough strength hardware platform to support,while traditional CPU has the bottleneck because of the von-Neumann architecture,its performance of computing has been unable to meet the demand of the required high performance computing,and the characteristics of high-performance GPU due to its high power consumption,but it cannot be used in embedded end mass.In view of the above problems,this study aims at efficiently understanding the structured information in remote sensing images.This paper based on the object detection model of oil palm tree,and we aim at the problems of low accuracy and low detection efficiency of high-resolution oil palm detection in deep learning,an effective and reliable solution is proposed from two aspects of algorithm optimization and heterogeneous hardware platform acceleration.Taking YOLOv3 object detection algorithm as an example,the optimization strategy of expanding feature selection and increasing multi-scale feature fusion is adopted to improve the detection accuracy of the algorithm for high-resolution oil palm.The result shows the high-resolution palm in Malaysia in this study in the area of about 55 square kilometers on the overall accuracy up to 94.53%,This is an improvement of 16 percent over existing tree identification methods,significantly reducing the confusion between oil palm and other vegetation.Then,the principle of deep learning algorithm and the complexity and parallelism of convolutional neural network are analyzed as well.Based on the high-performance heterogeneous platform,the high parallelism is realized by deploying a long parallel unit stream corresponding to the actual algorithm.The hardware design parameters of the convolutional neural network accelerator based on SIMD architecture and the effect of data multiplexing mode on scale scalability are studied and analyzed.An FPGA-based SIMD convolutional neural network accelerator with two-dimensional expansion is designed and implemented.An efficient convolution computing engine based on SIMD is designed by using the optimization strategy of weight shaping 8-bit quantization and computing core reuse.In addition,the input module is improved to speed up,and the input picture is sent to the input module in the form of write queue after dimensional change and vector-quantization processing,so as to improve the utilization of bus bandwidth.Experiments show that the performance of 1.4tops /s can be obtained on the heterogeneous hardware platform based on Intel Arria 10,and the performance is 7.51 times as high as that of the i9-9980 xe CPU,and the energy efficiency is 33.02 times as high as that of the i9-9980 xe CPU. |
PDF Full Text Request