【作者】 张毅；

【导师】 杨媛；

【作者基本信息】 西安理工大学 ， 微电子学与固体电子学， 2020， 硕士

【摘要】 视觉是人类获得外界各种重要信息的主要渠道。因此,视觉系统对人体生理系统来说非常重要,一旦受到损害将造成不可逆转的后果。随着科技的进步,视觉假体有望为盲人恢复一定程度的功能性视觉。而图像处理模块就是人工视觉假体系统中最重要的组成部分。因此对图像处理以及分类方面的研究具有重要意义。本文主要研究室内物体图像的分类算法和基于Zynq的硬件加速平台在人工视觉假体中的应用。在数据集方面,由于盲人的活动区域一般在室内,因此本文基于室内场景选取床,柜子,沙发等八类物体制作数据集。在算法方面,由于传统算法通常只采用某一种主要算法结构进行改进。本文则采用了一种将支持 向量机和卷积神经网络相结合的分类学习方法(CNN_nSVM)。多层卷积神经网络对于图像的特征提取效果较好,因此先用网络对图像的特征进行提取,最后一层采用SVM分类器替换CNN网络的分类输出层。在此基础之上,本文对支持向量机的分类核函数也进行了改进,将传统单一的核函数改进为组合核函数,使算法具有较低的损失值和较好的泛化能力。最后,在自建的室内物体数据集和官方数据集上对本文算法和相关文献算法进行测试。实验结果表明,在数据集MCIndoor20000上,本文算法分类精度比传统算法LBP+SVM、HOG+SVM分别高15.2%、9%,比相关文献算法 CNN+COSFIRE、Bachiri et al method 高 4.2%、6.1%。在自建数据集上,本文算法分类精度比传统算法LBP+SVM、HOG+SVM分别高20.1%、9.7%,比相关文献算法 CNN+COSFIRE、Bachiri et al method 高 3.2%、5.4%。本文完成了基于Zynq的室内物体分类算法硬件平台设计。详细介绍了各个模块的硬件设计方案,并且采用Verilog语言完成各个模块的设计,将改进后的SVM算法采用HLS封装成IP核以实现算法的硬件设计。同时利用Vivado Simulation仿真工具对设计的模块进行功能仿真,最后利用所搭建的Zynq硬件平台对算法模型进行验证。通过使用Xilinx集成的在线逻辑分析仪抓取关键点的数据进行验证,实验结果与仿真结果一致。最终实验结果表明,系统最高时钟为150MHz,硬件加速效果明显,在功耗方面比相关文献中低3.2-19.14W。更多还原

【Abstract】 Vision is the main channel for humans to obtain various important information from the outside world.Therefore,the visual system is very important to the human physiological system.Once damaged,it will cause irreversible consequences.With the advancement of technology,visual prostheses are expected to restore a certain degree of functional vision for blind people.The image processing module is the most important part of the artificial vision prosthesis system.Therefore,the research on image processing and classification is of great significance.This paper mainly studies the indoor object image classification algorithm and the application of Zynq-based hardware acceleration platform in artificial vision prosthesis.In terms of data sets,since the blind’s active area is generally indoors,this paper selects eight types of objects based on indoor scenes,such as beds,cabinets,sofas,etc.,to make data sets.In terms of algorithms,because traditional algorithms generally only use a certain main algorithm structure for improvement.In this paper,a classification learning method(CNN_nSVM)using support vector machine and convolutional neural network is proposed.The multi-layer convolutional neural network has a good effect on image feature extraction,so the network is used to extract the image features,and the last layer uses the SVM classifier to replace the classification output layer of the CNN network.On this basis,the classification kernel function of the support vector machine has also been improved,and the traditional single kernel function has been improved to a combined kernel function to make the algorithm have a lower loss value and better generalization ability.Finally,the algorithm of this paper and related literature algorithms are tested on the self-built indoor object data set and official data set.Experimental results show that,on the dataset MCIndoor20000,the classification accuracy of the algorithm in this paper is 15.2%and 9%higher than the traditional algorithms LBP+SVM and HOG+SVM respectively,and 4.2%and 6.1%higher than the related literature algorithms CNN+COSFIRE and Bachiri et al method..In the self-built data set,the classification accuracy of the algorithm in this paper is 20.1%and 9.7%higher than the traditional algorithms LBP+SVM and 3.2%and 5.4%higher than the related literature algorithms CNN+COSFIRE and Bachiri et al method.This paper completes the design of hardware platform for indoor object classification algorithm based on Zynq.The hardware design of each module is introduced in detail,and the design of each module is completed using Verilog language.The improved algorithm is packaged into an IP core using HLS to implement the hardware design of the algorithm.At the same time,the Vivado Simulation simulation tool was used to perform functional simulation on the designed modules,and finally the Zynq hardware platform was used to verify the algorithm model.By using Xilinx integrated online logic analyzer to grab the data of key points for verification,the results are consistent with the simulation results.The final experimental results show that the maximum clock of the system is 150MHz,the hardware acceleration effect is obvious,and the power consumption is 3.2-19.14W lower than that in the related literature.更多还原