FPGA Accelerator Design For CNN-based SAR Target Recognition

As a classification technique,Synthetic Aperture Radar(SAR)target recognition can extract features of the valid target from the interferential data including background clutter,which is widely used in the military attack and the intelligence acquisition.With the improvement of SAR image resolution,the difficulty of target recognition is gradually increasing.Hence,the research on quickly recognizing the target from massive SAR images with high accuracy becomes a hotspot.Recently,the Convolutional Neural Network(CNN)is applied to recognize SAR targets by many researchers due to its advantage of excellent feature extraction ability.However,with the increase of the CNN complexity,its performance is unsatisfactory on the Central Processing Unit(CPU)platform to accomplish large-scale calculations,especially for real-time target recognition.Although the Graphics Processing Unit(GPU)platform has outstanding acceleration performance,its power consumption is too large to meet the needs of engineering applications.Compared with CPU and GPU,the Field-Programmable Gate Array(FPGA)is regarded as the most appropriate hardware acceleration platform with many advantages such as low power consumption,the high degree of parallelism and flexible development.According to the above background,this paper focuses on the SAR target recognition method based on CNN,and meantime FPGA is used as the acceleration platform to design an efficient CNN accelerator of the SAR target recognition method.First,aiming to obtain a more accurate result in the SAR target recognition,the improved CNN,which is based on the fully convolutional neural network,is proposed to recognize SAR targets.At the same time,the Moving and Stationary Target Acquisition and Recognition(MSTAR)dataset is extended to avoid the problem of network overfitting.The experimental results including the analysis of CNN visualization,loss curves and accuracy curves demonstrate the effectiveness and practicability of the method presented in this paper.Second,for the purpose of solving the problem of multi-target recognition in SAR images,the fully convolutional neural network is introduced to improve the performance of CNN feature extractor in the Faster-RCNN.Since the SAR data is scarce,the multi-target dataset is extended and the VOC2007 standard dataset is used to train and test the network.The results under different conditions are analyzed,which confirms the validity of the strategy proposed in this paper.Last,in order to solve the problem of low efficiency on the CPU platform for SAR target recognition,the FPGA-ZYNQ Ultra Scale+MPSo C is used as the hardware acceleration platform according to the parallel architecture of the improved CNN.Each module of the accelerator is designed by using Verilog or C language with the collaborative method of software and hardware.In order to further improve the performance of acceleration,several optimal designs are investigated in the paper.Firstly,the convolutional vector processing unit,which is based on the 2D-convolutional unit,is designed with the degree of parallelism equal to 16,so that the efficiency of the convolutional module is improved.Secondly,the dual DDR cache structure is designed to improve data cache efficiency.Thirdly,the Softmax classifier is designed by using the C language to improve the classification accuracy.In the meantime,the minimum delay is achieved by four optimization methods presented in this paper.Then,the performance of the FPGA accelerator is analyzed minutely in terms of the acceleration performance,recognition efficiency,resource utilization and power consumption.The FPGA accelerator is finally compared with other acceleration schemes to emphasize the efficiency and advantages.