Design And Implementation Of Sar Object Detection System Based On A New Independent Intelligent Computing Platform

Different from the imaging method of regular optical imagery,synthetic aperture radar(SAR)uses electromagnetic pulses and terrain scattering characteristics for its image formation.This causes difficulties in parsing the SAR images.Furthermore,with the continuous development of space technologies worldwide,the number of satellites loaded with synthetic aperture radar is increasing each year;therefore,more difficulties arose in processing the exponentially increased SAR image data.How to utilize deep learning technology to quickly extract useful information from the mass amount of SAR images became a worldwide research focus.The advancement of software technologies needs supports from hardware platforms;however,due to the constant changes of the international context in recent years,the hardware supplies to China often encounter serious technology blockades.To improve the technology independence of China,there are many domestic companies that are developing intelligent processors,which are specifically for the applications of artificial intelligence.The deployment of SAR image detection on such independent intelligent computing platforms,therefore,has significant importance to China's national interests and securities.This paper explores the application of convolutional neural networks for SAR image detection on China's independent intelligent computing platform.First of all,due to the special application scenarios of SAR image detection,its hardware platforms have a high requirement for resource efficiency and model size;therefore,the traditional image detection models cannot be deployed on such resource-constrained computing platforms.To resolve this difficulty,this paper uses the depthwise separable convolution modules to re-design network structures and proposes a new single-object SAR image classification model based on a lightweight neural network structure.To avoid the problem of information loss during training,this paper proposes a new activation function – SRe LU activation.Meanwhile,aiming to mitigate the network overfitting caused by the dataset sparsity,this paper augments the original dataset,which improves the generalization ability of the network.Secondly,aiming at the problem of crowded scene detection,this pape r proposes a new neural-network based SAR crowded scene detection algorithm,which uses the SAR single-object classification network as its network backbone to reduce the model size.This proposed model uses the low-level feature fusion method to improve its performance for small object detections.Meanwhile,it also introduces an RFB module as an additional feature extraction layer to improve its feature extraction ability.Thirdly,to satisfy the computation precision requirement of independent intelligent computing platforms,this paper quantifies the trained models.Then the model is successfully deployed on an independent intelligent computing platform.According to the operation characteristics of the independent intelligent computing platform,this paper further optimizes the object detection model using AI programming language,which improves the model's calculation efficiency.To explore the model's best acceleration performance on the independent intelligent computing platform,this paper tests different combinations of model structures and operating environments.By setting effective evaluation methods,this paper reviews the performance of all different combinations.Finally,based on the foregoing models,this paper designs and realizes SAR object detection system based on a new independent intelligent computing platform.