| With the continuous development of social economy,China has become a large country of production and manufacturing,producing a large amount of food,medicine,etc.every year.Manufacturers will spray-print production dates,lot numbers and other information on their goods,which has become an important way for people to pay attention to food safety.At present,a few companies use the manual quality of the spray code character detection,but the manual detection of low efficiency and high labor costs,so most companies use the machine vision method instead of manual to achieve intelligent automatic identification of the spray code,but the traditional machine vision in the complex background of the spray code recognition accuracy is low,for different packaging needs to design a different method,poor generalization.The IPCs running the algorithm at the same time are large and inconvenient to use.As hardware devices continue to evolve,it has become possible to run deep learning models on embedded devices.This paper focuses on the application of OCR techniques in deep learning to the field of jet code character recognition,with some improvements to improve the accuracy of the model for jet code recognition,in addition to deploying the used model on Jetson devices to make the recognition method more portable.The main research work is as follows.Firstly,an experimental platform for the recognition of spray-coded characters is built,and a detailed analysis and selection of cameras,lenses,light sources and edge computing platforms among the hardware devices are carried out.It is mainly used to capture the spray code images and verify the feasibility of the proposed algorithm.Secondly,the DBNet-based jet code character detection model is designed.In order to improve the comprehensive performance of the original model for jet code localization,CBAM(Convolutional Block Attention Module)module is added to the backbone network;meanwhile,in order to use the model to have higher inference efficiency on Jetson platform,the number of parameters of separable convolution and standard convolution is compared,and it is decided to use Mobile Net V3 network constitutes the backbone network of DBNet.Thirdly,the DBNet-based Code-spurting character detection model is designed.In order to improve the comprehensive performance of the original model for jet code localization,CBAM(Convolutional Block Attention Module)module is added to the backbone network;meanwhile,in order to use the model to have higher inference efficiency on Jetson platform,the number of parameters of separable convolution and standard convolution is compared,and it is decided to use Mobile Net V3 network constitutes the backbone network of DBNet.Finally,the deployment of the model on Jetson devices and inference optimization are completed.The model is deployed to the Jetson platform and the model is optimized for inference using Tensor RT as well as quantization to improve the inference speed of the model.Since Tensor RT acceleration can have some impact on model performance and restrict the hardware platform.To further investigate the model embedded deployment,this paper also employs the neural network compiler TVM for inference acceleration.There is a large improvement in inference speed compared to the original model. |
