Research And Implementation Of Image Classification And Recognition Technology Based On PYNQ

As an important part of artificial intelligence,image classification and recognition has broad application prospects and important research significance.Convolutional neural network model is a very widely used technical means in image classification and recognition and has been deployed on CPU and GPU platforms,but the defects of high power consumption make CPU and GPU unable to be applied in current embedded mobile terminal scenarios.In recent years,new research hotspots have focused on how to implement image classification and recognition systems with low power consumption.This thesis proposes an image classification and recognition system solution based on Xilinx's PYNQ-Z2(Python productivity for Zynq)embedded development platform,and studies the implementation method based on convolutional neural network and ARM + FPGA heterogeneous system.The software and hardware collaborative design ideas are used to plan software and hardware work tasks,formulate system performance indicators,and finally build a PYNQ image classification and recognition system that can recognize different data sets by editing the upper computer program to read the characteristic parameters of different data sets.Not only significantly improves the generality of the system,but also significantly reduces hardware power consumption on the premise of meeting the classification recognition function.The main research work in the thesis includes the following aspects:First,the overall system is analyzed,the functional modules are clarified through software and hardware collaborative design ideas,and task division is completed.In the processing system(Process System,PS)part,the Jupyter Notebook platform is used to implement the reading of binary characteristic parameters of the host computer program and control of the hardware based on Python.In the programmable logic(Programmable Logic,PL)part,the module design and system path construction of the convolutional neural network are realized.After understanding the basic principles of the convolutional neural network algorithm,build the convolutional neural network model on the computer-side MNIST dataset and CIFAR-10 dataset to be tested,complete the training verification and finally get the accuracy of the MNIST model and CIFAR-10 model They are 99.06% and 62.25%,respectively,and then the feature parameter extraction function is designed to complete the weight and paranoid parameter extraction and format conversion,and convert to a binary format that can be read by the hardware platform.Then use Xilinx VIVADO HLS(High Level Synthesis,HLS)design tool to design and implement a custom IP core module for the convolutional neural network in the image classification and recognition system,including the convolutional layer IP core and the largest pooling layer IP core.Using the HLS tool The optimized instructions in the combination are compared and tested,and finally an IP core module that meets the design goals is synthesized.After the design of the custom IP core is completed,the IP core module and the ZYNQ module are mainly used to implement the path construction of the overall system.After the verification is completed,the control is called by the host computer program in Jupyter Notebook.Finally,the design of the driver and the host computer is completed,and the system is tested for functions and performance.The test shows that the system can achieve normal classification on the recognition of the MNIST and CIFAR-10 datasets,and the system power consumption parameter is 1.54 W.The test results show that only through different feature parameter files and editing the upper computer program can achieve the universal design goals of the system,the power consumption of the system is far lower than the power consumption of traditional platforms such as CPU,which proves the PYNQ-based image classification recognition system Feasibility.