SAR Image Target Detection Algorithm Based On Deep Learning Accelerates Research

With the rapid development of artificial intelligence,the application of deep learning technology in the Internet of Things industry is booming.SAR satellite image detection,pedestrian detection,autonomous driving and so on are emerging.However,in order to adapt to the high accuracy of the detection effect target detection network layer number is designed to be deeper and deeper,at the same time the resulting computational volume has become larger.We know that every ring from the beginning of the design to the actual deployment of the target detection algorithm is critical,not only at the expense of a significant increase in computational volume to improve the accuracy of the detection,but also to take into account the difficulties encountered in deploying the target detection algorithm to real-life scenarios.Embedded devices for large-scale applications in real life tend to have limited budgets(low computing power),such as NVIDIA Jetson Nano,where low power consumption can be powered by USB for only 5w,and NVIDIA Jetson Xavier NX,which has a maximum power of 15 w,for example,they cannot be compared with server-side devices with high computing power.Therefore,the design of a qualified and realistic target detection network can not be in order to pursue high accuracy and ignore its practical application ability.So how to optimize the design of the network without losing a lot of precision,so that the model reasoning speed is improved is the main research direction of this paper.Based on the image detection of SARS ships,the results of the main innovation work and research are summarized as follows:(1)In order to solve the difficulty of real-time inference target detection model on embedded platform with limited computing power,this paper puts forward the acceleration scheme of target detection algorithm based on deep learning.First,the feature extraction network is replaced by a target detection model based on YOLOV3 original high computation.Replace complex feature extraction networks with lighter MobileNetV2 with some accuracy.Secondly,the complex target detection model is structured and pruned,and the number of parameters of 30% and 50% of the original model is cut,respectively,so that some redundant parameters can be removed to obtain better reasoning speed.In the end,compared with the original SAR image detection model,the average loss of control accuracy was about 1.47%,and the average speed could be increased by 40.1%.The SAR image detection model with optimization scheme improves speed while ensuring accuracy.(2)In order to solve the difficulty of training the completion target detection algorithm from the server to applying it on a small intelligent platform,a complete set of executable deployment scheme is proposed.Using the TensorRT forward reasoning acceleration scheme provided by NVIDIA,the paper uses TensorRT to fuse the structural similarity of the original target detection model,and accelerates the effect by reducing the call to the CUDA core,while the int 8 low-bit quantization method of TensorRT is used to further speed up the target detection model.The open neural network switching format(ONNX)can be used between different frameworks for model-to-model conversion,and after converting to a Model supported by TensorRT,it can be reasoned directly from the original training framework.The experiment resulted in a complete process from training on the server to complete the target detection algorithm to deploying it on a small intelligent platform.