基于卷积神经网络的单图像去雾模型硬件重构加速方法  

Hardware reconstruction acceleration method of convolutional neural network-based single image defogging model

作　　者：王官军 简春莲 向强[1] WANG Guanjun;JIAN Chunlian;XIANG Qiang(College of Electronic and Information,Southwest Minzu University,Chengdu Sichuan 610041,China;Chengdu Powerview Science and Technology Company Limited,Chengdu Sichuan 610046,China)

机构地区：[1]西南民族大学电子信息学院,成都610041 [2]成都动力视讯科技股份有限公司,成都610046

出　　处：《计算机应用》2022年第10期3184-3190,共7页journal of Computer Applications

基　　金：四川省科技厅重点研发项目(22ZDYF3125);西南民族大学研究生创新型科研项目(CX2021SZ46)。

摘　　要：针对基于卷积神经网络(CNN)的单图像去雾模型在移动/嵌入式端部署难,不易用做实时视频去雾的问题,提出一种基于Zynq片上系统(SoC)的去雾模型硬件重构加速方法。首先,提出量化-反量化算法,对两个代表去雾模型进行量化;其次,基于视频流存储器架构和软硬件协同、流水线等技术以及高级综合(HLS)工具,对量化后的去雾模型硬件重构并生成具有高性能扩展总线接口(AXI4)的硬件IP核。实验结果表明,在保证去雾效果的前提下,可以实现模型参数从float32到int5(5 bit)的量化,从而节省约84.4%的存储空间;所生成硬件IP核的最高像素时钟频率为182 Mpixel/s,能够实现1080P@60 frame/s的视频去雾;单帧640×480的雾图去雾仅需2.4 ms,而片上功耗仅为2.25 W。这种生成带有标准总线接口的硬件IP核也便于跨平台移植和部署,从而可以扩大这类去雾模型的应用范围。Single image defogging model based on Convolutional Neural Network(CNN) was difficult to deploy on mobile/embedded system and used for real-time video defogging. To solve this problem, a method of hardware reconstruction and acceleration was proposed, based on Zynq System-on-Chip(SoC). First, a quantization-dequantization algorithm was proposed to perform quantization on two representative defogging models;second, a quantized defogging model was reconstructed and a hardware IP core with Advanced eXtensible Interface 4(AXI4) was generated, based on video stream memory architecture, hardware/software co-design, pipeline technology and High-Level Synthesis(HLS) tool. Experimental results show that the model parameters can be quantified from float32 to int5(5 bit) under premise of defogging performance, saving about 84. 4% of storage space;the highest pixel clock frequency of the generated hardware IP core is 182 Mpixel/s, which can achieve 1080P@60 frame/s video defogging;the hardware IP core processes a single hazy image with the resolution of 640 pixel × 480 pixel only in 2. 4 ms, and the on-chip power consumption is only 2. 25 W. This hardware IP core with AXI4 is also convenient for cross-platform migration and deployment, which can expand application scope of CNN-based single image defogging model.

关 键 词：去雾 视频直接存储器访问 模型量化 模型重构 硬件IP核 高级综合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]