| Part of the production process in the processing of industrial products,such as shrinkage and solidification,will cause defects such as scratches and wear on the surface of the product.In traditional production processes,manual quality inspection is usually used to judge product quality.This inspection method not only consumes labor costs,but also cannot accurately detect product quality.With the improvement of the intelligence level of the manufacturing industry,the defect detection system based on the deep learning algorithm gradually replaces the manual quality inspection,which improves the defect detection efficiency and labor productivity.However,most of these complex deep learning algorithms need to be deployed on cloud servers for training,which cannot meet the needs of offline execution and are difficult to deploy into embedded platforms.In response to these problems,this dissertation designs a surface defect detection algorithm training platform for industrial products,and conducts lightweight research on the detection algorithm to realize the local training of the model.The specific work is as follows:(1)Considering the factory environment and functional requirements,this dissertation proposes a design scheme of a localized training platform for surface defect detection algorithms.The Inception v4 model is used to judge whether the product is qualified and the product defects are identified based on the YOLO v5 model,and a localized training platform that integrates the two models is designed.The platform can use product images to train detection models,supplemented by database tools to store training task data,and design experiments to perform functional tests on the platform with circuit board product data sets as an example.(2)Take the model accuracy and model size as the primary conditions,carry out lightweight research on the detection model,and design two network structure optimization strategies.One is to introduce a scaling factor in the network training process to identify unimportant channels to trim the model;the other is to quantify the network parameters during the training process and convert 32-bit floating point data to 16-bit integer data.to save storage space.And experiments are carried out on the data set of circuit board products collected by ourselves,and the results prove that these two optimization strategies can compress the model size.(3)In order to verify the feasibility of the model’s lightweight,taking product classification as an example,an application example of the surface defect detection system for industrial products is designed.Supported by the general embedded computing architecture,the detection model parameters are converted into general embedded engineering components,which simplifies the application process of product classification algorithms in embedded artificial intelligence.Users only need simple operations to complete product classification detection at the terminal.In this dissertation,a local training platform based on the deep learning surface defect detection algorithm is designed for industrial products,and the lightweight research of the detection algorithm is carried out to reduce the detection cost and improve the detection efficiency,which has certain engineering application prospects. |
