Implementation Of Low-power FPGA Based On LeNet Network Target Recognition

With the rapid development of convolutional neural networks,the size of input images is getting larger and larger,the number of network structures is increasing,and the complexity of the structure is getting higher and higher,and the target recognition network parameters based on hardware are increasing,it is more and more difficult to design a balance in speed,area and power consumption.Therefore,this paper proposes a new LeNet network structure,the recognition accuracy can reach more than 97%.Compared with the design in literature [37],the design of this paper can save 49% of LUT(Look-Up-Table)resources,83% of FF(Flip Flop)resources and 91% of DSP(Digital Signal Processor)resources under 50 MHz clock.At the same time,the power consumption is only 0.865 W,which realizes the relative balance of speed,resources and power consumption of the neural network.The LeNet network structure designed in this paper consists of 7 layers,including 2convolutional layers,2 pooling layers,2 fully connected layers and a SoftMax layer.This paper analyzes the various modules of the LeNet network structure in detail.Among them,the convolutional layer and the pooling layer are the structures for extracting image feature data,the fully connected layer is to comprehensively summarize the extracted feature data,the SoftMax layer is the output layer,and the output Recognition result and corresponding accuracy rate.This paper improves the accuracy of the network by improving the training and weighting method of the network at the algorithm level.At the same time,in the hardware design,the most complex convolutional layer in the calculation process is improved and the calculation structure of the SoftMax layer is optimized.Through a detailed analysis of the difficulty of convolution calculation and the study of traditional design architecture,a targeted optimization plan is proposed.In the process of calculating the exponential function of the SoftMax layer,a method of combining approximate calculation and look-up table is proposed,which converts the complex calculation into a simple look-up table process,which simplifies the calculation process and reduces the resource occupation.The optimization of resource occupation and power consumption is realized by optimizing the complex structure.The design process of this paper is mainly divided into two parts.One part is the construction of the LeNet network training model,which is mainly responsible for the training of the MNIST data set from the algorithm level and the quantification and extraction of weight data.This part is designed in the Visual Studio 2017 and Python 3.6 environment.To complete the construction and training of the training model by using the TensorFlow library;The other part is the design of LeNet network hardware model.It mainly uses Verilog language to design the various functional structures of the network,and performs functional verification by building Testbench.Finally,the verified code is encapsulated into IP cores and added to the LeNet network system.After comprehensive implementation,the bitstream file is downloaded to the NEXYS VIDEO development board for verification.This part of the design uses Vivado as the development software for code writing and simulation testing.