Image Feature Point Extraction Based On Neural Network Is Implemented On FPGA

With the continuous development of computer vision technology,image recognition technology is widely used in the fields of automatic driving,nautical aviation,medical treatment,etc.In the field of computer vision applications,feature points play an important role.At present,most image processing is based on software,but the huge amount of computation will cause the whole system to be inferior in performance and speed.Although there is a dedicated graphics processor(GPU),it is also at the expense of power consumption.of.In the traditional image feature point extraction algorithm,feature point extraction such as SIFT,SURF,etc.are all dependent on the characteristics of manual design,and the characteristics of manual design are too cumbersome.In view of some of the above problems,this paper mainly studies the method of neural network image feature point extraction based on FPGA.In the comprehensive analysis and comparison of the advantages and disadvantages of programmable logic gate array(FPGA)and CPU and GPU,this paper finally adopts the "ARM+FPGA" design method,using Xilinx's PYNQ FPGA platform,which includes Programmable Logic(PL)and Processing System(PS)end,through the combination of software and hardware,reasonable allocation of hardware and software tasks,and finally achieve image feature point extraction,which not only increases the flexibility of the entire system operation.Sexuality,and greatly reduce power consumption while maintaining performance.The main contents of this paper include the following aspects:Firstly,the function of the whole system is comprehensively analyzed,and then the system is divided into reasonable software and hardware tasks.On the PL side,the related calculation of convolutional network is realized.On the PS side,the hardware control and feature point extraction related algorithm are implemented in combination with PythonAfter deep understanding of the algorithm of the convolutional neural network,the Xilinx Vivado HLS(High Level Synthesis)tool is used to design and implement the IP core of each layer of the convolutional neural network in a customized way.In the process of optimization,the HLS tool is utilized.In the optimization instructions,it is necessary to analyze and compare various optimization schemes,and finally optimize the design of the IP core in terms of speed and area.After custom designing the IP core,add these IP cores to the Xilinx Vivado tool,center the custom IP core,and combine the IP cores that come with the Vivado tools toimplement the layout of the IP core.After the design verification,the final Generate bit files and tcl files,and control these files through the PS side call.Finally,the paper tests and analyzes the whole system design.The test results show that the system can extract more useful feature points in image feature point extraction,and rotate,zoom,change illumination and seasonal changes in the image.Other features have better performance,and the matching effect is better than the SIFT matching.The system takes about 3 seconds to extract a 640x480 image,and the performance is about the same as that of the GPU.The FPGA is about 2W in power consumption,which is far lower than the power consumption of the GPU.