| Aiming at the low detection efficiency of modulation recognition algorithm based on deep learning at present stage,this thesis proposes an efficient modulation recognition algorithm named RadioFSDet detection algorithm combining with the object detection YOLOv4 algorithm in the field of computer vision.The general process of the algorithm is to convert the one-dimensional VHF frequency spectrum data into two-dimensional image data,and then use the object detection algorithm in deep learning to detect the modulation signal in the image,mainly detect the modulation category of the signal,and roughly determine the central frequency of the signal.Compared with the current mainstream modulation recognition algorithm based on deep learning,the algorithm has high detection speed.In addition,in order to further improve the inference speed of the RadioFSDet algorithm,this thesis combined with the achievements in the field of model lightweight and model quantization at the present stage,aiming at the RadioFSDet algorithm,the following work is done:The backbone of object detection model YOLOv4 is improved in lightweight.In this thesis,a more efficient lightweight backbone network CS-LCNet is designed,and combined with Neck network of YOLOv4 to construct a RadioFSNet object detection model.The experimental data shows that while ensuring the accuracy,the inference speed of RadioFSNet is much higher than YOLOv4,and model inference delay on the x86 CPU reaches 73.6ms.The post-training quantization scheme is used to quantize RadioFSNet,but it is found that the accuracy loss of RadioFSNet quantized by the scheme is serious.Therefore,this thesis further analyzes the causes of serious loss of accuracy in the process of quantization,and combines the algorithms of relevant literature to improve the post training quantization scheme.Experimental results show that the model accuracy of RadioFSNet is greatly improved after quantization of the improved scheme.The overall detection process of RadioFSDet is summarized,and different inference frameworks are selected for different hardware platforms to complete the deployment of RadioFSDet algorithm.Specifically,for the x86 CPU hardware platform,the Open VINO framework is used for deployment,the model inference delay is reduced to 50.5ms.For the hardware platform of the mobile device,the NCNN framework is selected for deployment.The INT8 model inference delay on the mobile device is reduced to 44.76 ms. |
