Research On Infrared Image Intelligent Recognition Algorithm And Hardware

In recent years,with the continuous development of neural network technology,its application field has become more and more extensive,especially the convolutional neural network has improved the problem of traditional algorithms requiring manual extraction of features,and has achieved outstanding results in the field of image recognition.Infrared images,as a supplement to traditional visible light images,have irreplaceable advantages in certain specific scenarios,such as nighttime and heavy fog.Due to the generally low resolution of infrared images and the lack of details,how to perform high-speed processing of infrared images has also become a hot research topic.At the same time,due to the limitations of the hardware resources of traditional terminals,many practical application scenarios of algorithms that perform well on ordinary computers are limited on the terminal.How to run these algorithms on a small,low-power platform also requires us to study.This article conducts research from the following aspects:Firstly,the mathematical model of artificial neural network,the process of forward propagation and back propagation,and how to train the network are analyzed in detail.And pointed out the shortcomings,and then proposed how the convolutional neural network solves this shortcoming,and introduced several typical structures of convolutional neural networks and how they operate.Then optimize and improve according to the classic LeNet-5 network model,and propose a neural network model that can be used for infrared image recognition.The neural network and data set were built and trained through the pytorch network framework,and the 96%recognition accuracy rate was finally achieved on the test set.Then,for the application scenarios with limited resources such as terminals,several schemes for network model compression are proposed,and the compression schemes of convolutional layer quantization and fully connected layer pruning are selected and applied on the infrared image recognition network.94%,2 percentage points lower than the original network,but the overall size of the compressed network is 35%of the original network,which achieves a reduction in recognition accuracy within an acceptable range,and greatly compresses the network size,making it in the hardware More flexible in implementation.Considering the large number of parallel features of convolution operations,a design scheme of FPGA-based convolutional neural network accelerator is proposed.Through Vivado HLS high-level synthesis,using C/C++ language description algorithm and then converted to RTL level implementation.At the same time,a variety of optimization instructions are used,including pipeline optimization,full loop expansion,data type optimization,and other methods to optimize the convolutional layer,pooling layer,and fully connected layer at the hardware level,and export them in the form of IP cores Then use the Vivado integrated development environment to complete the construction of the final system and generate the corresponding bitstream file.On the PYNQ-Z2 platform,the bit stream file was burned and the system function was tested.The recognition accuracy of the infrared recognition network implemented by hardware was 93%,and the performance and power consumption were compared with the traditional CPU/GPU platform On the comparison.The calculation speed is 4.3 times faster than the traditional CPU platform,but 1.2 times slower than the GPU platform;but in terms of power consumption,the CPU platform is 37.6 times that of the GPU platform,while the GPU platform is 61.6 times.The test results show that the system achieves the expected performance indicators and meets the design requirements.