Implementation Of Printed IC Chip Character Recognition System Based On Heterogeneous Platform

With the increase in number and speed of IC chip production,quality control has become more and more important.One of the tasks of quality control is to detect whether the printed characters of IC chips are misprinted.Since the printed characters of an IC chip contain the name,performance and function information of the chip,misprints which are not detected in time may lead to serious errors later on.Considering that the string on the chip is very small and the number are very large,the traditional detection methods such as manual detection and template matching cannot complete the detection work with high efficiency and reliability.Aiming at the specific requirements of IC chip printed character recognition,this thesis built a printed IC chip character recognition system based on ZYNQ which is an ARM+FPGA heterogeneous platform.Image processing algorithms in the system mainly include grayscale,binarization,image filtering,image morphology processing and character location.The character recognition algorithmis realized by convolutional neural network.Besides,hardware acceleration of the above algorithm greatly accelerates the running speed of the system.The main work of this thesis is as follows:(1)Based on the investigation of the software and hardware co-processing technology and the image classification technology,ZYNQ was selected as the hardware core of the system,and the convolutional neural network has used as the image classification algorithm.CMOS camera used to realize high-speed image acquisition.ZYNQ used to process and recognize the collected images.At last,the system output results of recognition.The system uses AXI bus to connect each module,and uses VDMA to realize the efficient and orderly transmission of data.(2)Location algorithm based on connected domain has used to locate the chip from the image and screen out the chip region conforming to the standard.The chip image needs to deal with grayscale,binarization,image filtering,image morphology processing and other pre-processing steps after location.At last,character segmentation has carried out on the preprocessed image.The results of preprocessing show that the effective information of the image is enhanced after preprocessing.(3)The convolutional neural network classification architecture has adopted in the system.After analyzing the parallelism of the convolutional neural network Lenet-5,the convolutional neural network module accelerated by FPGA has been designed.The acceleration of convolutional neural network is realized by intralayer parallel optimization method,matrix multiplication optimization method and the optimization method which using fixed point number instead of float point number.The optimization method of matrix multiplication is mainly realized by pipelining and unrolling.By comparing the number of cycles with the number of resource consumption,the specific strategy is determined.(4)The hardware system has been built with Vivado.And the embedded system of printed IC chip character recognition has been integrated.The Linux operating system has been transplanted,including some drivers of CNN IP and other IP.The scheduler has been writen with SDK and optimized for data transmission.(5)Through the experimental test,the image processing module of this system is similar to the image processing effect realized by ARM.And its speed is 11.49 times faster that of the latter.The character recognition module based on CNN also achieves 7.69 times faster of acceleration when the recognition rate reaches 98.68%.Besides,it takes 8.21 s to recognize 10000 images of 28*28*1.The power consumption of the system is 2.275 W,which achieves the predetermined purpose of low power consumption and high performance.