Design And Implementation Of FPGA-based Convolutional Neural Network Systems

Autism is a disease associated with the nervous system.Patients often have language and social disorders in their daily lives,or they have the same repetition of the same movement.The patient is diagnosed at an early age and receives early intervention,which results in better treatment.However,there is currently a problem lacking experienced teachers in autism education.The use of artificial intelligence,computer vision,robotics and other information technology to assist teachers in assessing the teaching status of students can reduce the pressure of teachers.There are several requirements for applying deep learning algorithms to autism teaching scenarios.In the classroom teaching,the teacher hopes to obtain timely analysis data of the classroom teaching video so that the teaching plan can be adjusted at any time;Outside the classroom,parents hope that the intervention treatment for autistic children in the family can also get help from information technology;In order to protect student privacy,classroom videos should not flow out of the classroom scene.Based on the above requirements,this paper believes that it is necessary to deploy a system that supports deep learning applications in autism teaching scenarios.A set of FPGAbased convolutional neural network acceleration system was designed to support the application of deep learning algorithms in autism teaching scenarios.The system is divided into three modules,namely video acquisition module,convolutional neural network accelerator and data output module.The design and implementation of convolutional neural network accelerator is the main work content of this paper.In this paper,the HLS technology is used to develop the convolutional computing unit and the pooled computing unit of the convolutional neural network.The convolutional neural network accelerator is implemented on the Zybo development board and the KCU1500 development board through the development model of software and hardware co-design.The convolutional neural network acceleration system constructs a convolutional neural network model through application software,and uses the convolutional computing unit and the pooled computing unit in the FPGA to perform hardware acceleration on the convolutional neural network algorithm.In this paper,the function evaluation of the convolutional neural network acceleration system is carried out,and it is verified that the modules of the convolutional neural network acceleration system can operate normally in the autism classroom teaching scene.This paper also evaluates the performance of the convolutional neural network accelerator.By comparing the time spent running the same network,the Zybo development board runs a convolutional neural network accelerator to achieve 25 times performance acceleration of the dual-core Cortex A9 ARM processor.For,KCU1500 development board,the performance of the Intel i7-7700 processor is 10 times faster,which proves that the convolutional neural network acceleration system designed in this paper can provide the computational acceleration of the convolutional neural network in the autism teaching scene.