An Online Inspection Method For Packaging Defects Of Emulsion Explosives Based On Machine Vision

The identification and classification technology of emulsion explosive packaging defects is an important part of the production process of civilian explosives.This technology has the significance of industry development and scientific research and entrepreneurship.How to accurately and quickly identify the defects of emulsion explosive packaging is a major factor in the current civil explosive industry.Traditional artificial pipeline detection and traditional image processing technology are not only inefficient but also extremely unsafe.This thesis proposes an on-line detection system for emulsion defect based on machine vision,which has great significance and value for the identification of emulsion explosive defects.Firstly,a new algorithm for detecting defects in emulsion explosives based on convolutional neural network is proposed.The optimization process is reflected in three aspects:(1)A new flexible learning rate correction algorithm is proposed.The mutation rate is used to judge the learning rate instead of the fixed learning rate in the past.Dynamically calculate the adaptive weight coefficients of each stage relative to the current network according to the mutation factor obtained during training;(2)A threestage processing method is adopted for the convolutional neural network training process to achieve the purpose of flexible learning;(3)Adopt Leaky-Relu function instead of traditional sigmoid function as the activation function,to achieve better activation effect by adjusting the suppression factor.Using the BSDS500 data set as the experimental data set,the analysis of the results optimizes about 90% of the network training time and improves the recognition accuracy by about 20%.Furthermore,an image pre-processing scheme before defect detection is proposed,and a pre-processing scheme is established for the problems of image capacity,noise,adhesion,and unclear boundaries.After processing the solution,the image capacity is reduced by nearly 20 times and only the boundary information is left.The medicine roll image with only the boundary information is subjected to defect detection by the CNN algorithm based on the flexible learning rate,which greatly reduces the pressure of defect detection and improves Accuracy and speed of defect detection.The experiment uses 2000 pictures of medicine rolls as the data set,the detection speed is 2.3 minutes,and the defective detection rate reaches 98%.Finally,a medicine roll image acquisition module is designed,and the Raspberry Pi is used to control the light source and the camera at the same time.The three are composed of image acquisition units,and multiple units are composed of image acquisition modules.Control the light source and camera at any time,and realize the processing of the same batch of image data at the same time.Developed a visual drug roll defect detection platform with artificial intelligence characteristics.The system has an intuitive data representation and a simple operation process.The system developed by the project is implemented on site and put into the production environment.Through experiments and on-site environmental tests,the development system has reached industrial application standards.