Deep Learning Model Acceleration And Embedded System Implementation

Deep learning image processing technology is one of the most concerned technologies in academia and industry.However,the general deep learning model cannot be directly deployed to embedded platforms due to the long computation time,especially cannot be applied to realtime response scenarios.How to improve the speed of model inference while ensuring the accuracy and reliability is of great significance to the wide application of deep learning image processing technology.This thesis makes a detailed and in-depth study on the method of model inference acceleration.In order to solve the problem that the inference of deep learning model takes too long on embedded devices,based on the traditional data-driven channel importance evaluation method,a deep learning model channel clipping method based on grey correlation analysis is proposed.Based on grey correlation analysis method,in turn,will each channel as a reference sequence calculation for each channel layer relative to the other channel of average correlation degree of quantitative values,the importance of the channel and the value is inversely proportional to the size of the quantitative value,the greater the reference sequence extracted features and other channels,the more similar reference sequences corresponding convolution kernels channel important degree is lower,give priority to cut in the cutting process.Experiments show that this method can reduce the inference time of VGG model on embedded devices from 266 ms to 95 ms,and reduce the precision by only 1.9%.Aiming at the problem of accelerating upper limit of model compression by channel clipping,an optimal quantization boundary selection method based on KL divergence is proposed on the basis of traditional linear quantization,and the model weight parameters are quantized by ADMM.The good ability of KL divergence to represent the information difference between two distributions was used to balance the overall numerical representation width and the accuracy of a single numerical representation.The information loss before and after the quantization was selected as the optimal quantization threshold to maintain the quantified model accuracy to the greatest extent.ADMM method is used to model the weight at the same time quantitative mapping problem is transformed into optimization problem,by calculating the minimum parameters before and after quantitative difference of L2 norm as loss value,using the ADMM alternating update function caused by coefficients quantifying quantitative coefficient and contain quantitative values,until finally the loss value towards convergence when the corresponding quantitative factor is the best quantized coefficients.Experiments show that this method can further reduce the single reasoning time from 266 ms to 110 ms,and only reduce the precision by 4.6%.In order to build an embedded device system based on deep learning wood defect detection algorithm,Nvidia Jetson TX2 embedded computing board card and DALSA linear array camera are used as hardware,and YOLO V3 algorithm using MobileNet as the backbone network is used as detection algorithm.Real-time defect detection of wood was carried out in simulated real-time production environment.Experiments show that after model channel clipping,the inference time is reduced from 92 ms to 57 ms,while the mAP precision of 0.91 is guaranteed.After the model was further quantified with 8 bits,the detection time was only 25 ms when the accuracy mAP was 0.85.The mAP of 0.85 and the detection speed of 40 frames per second can meet the accuracy and speed requirements of flaw detection in the wood production site.The results verify that the proposed method is feasible to deploy high speed wood defect detection system for embedded equipment.