Design And Implementation Of Image Classification System Based On FPGA

In recent years,image classification has a wide range of application scenarios,including satellite image detection,medical image processing,geomorphic image recognition,face recognition,environmental monitoring and so on.At the same time,the demand for the application and deployment of image classification algorithms in hardware terminals is increasing day by day.How to transplant these algorithms to hardware terminals and optimize their performance has become a hotspot and difficult area on research of image classification.To this end,this dissertation proposes an FPGA-based image classification algorithm and transplants as well as accelerating the algorithm on Xilinx's Zynq-7000 series platform to verify the effectiveness of the algorithm.The main contents of this article include:First,this dissertation determines the design platform of the image classification system based on the implementability of the algorithm transplantation,puts forward the algorithm flow of the overall system,and determines the software and hardware architecture as well as the design scheme of each component of the system according to the design rules of So PC system.Secondly,the algorithm principle and design idea of image classification method based on manual features are studied in this dissertation.In terms of manual features,this dissertation selects SIFT features as manual features,and discusses the construction schemes of two SIFT feature libraries,the bag of words model and VLAD.At the same time,combined with the characteristics of the hardware platform,the design scheme of the algorithm transplanted on FPGA is proposed,and the optimization scheme of the algorithm in the feature point detection,gradient calculation,feature point description,etc.is given.Then,the algorithm principle and design idea of image classification method based on depth features are studied in this dissertation.In terms of depth features,CNN features is selected as depth features in this dissertation.In order to obtain the respective advantages of deep convolutional features and manual features,and improve the image's feature expression ability to obtain a better image classification effect,the two features are combined by this dissertation using series weighted fusion.At the same time,combined with the characteristics of the hardware platform,the design of transplanting the algorithm to FPGA is proposed,and the optimization scheme of the algorithm in parallel development of convolution and partial and cumulative calculation is given.Finally,this dissertation deploys the improved algorithm based on hardware to FPGA and gives the design method as well as implementation scheme of each module in detail.The implementation of digital design adopts the methods of spatial parallelism and pipeline design,which makes full use of the parallelism of FPGA and improves the operating efficiency of the algorithm.Experimental results show the hardware acceleration effect of FPGA is significant,and that the image classification system designed in this dissertation has real-time performance and achieves a high accuracy at the same time.