Medicine Glass Bottle Defect Detection Based On Machine Vision

Glass containers are widely used in medicine,chemistry and other fields.With the development of industrial automation,manual inspection of glass bottle defects has been unable to meet the demand.Using machine vision instead of human eyes to complete defect detection has become an inevitable trend in glass bottle defect detection.At present,domestic glass bottle inspection equipment based on machine vision is not as advanced as visual inspection equipment in developed countries such as Germany and America.However,due to technical confidentiality and other reasons,foreign machine vision equipment is very expensive,Leading to the limited industrial application of such equipment.This paper designs a method for glass bottle defect detection and defect range determination based on image processing algorithm.Designing the light source illumination method and experimental platform according to the production environment of the glassware and the light guiding property and refractive index of the glass material.Pretreatment the acquired glass bottle defect image,including image pretreatment steps such as remove image noise,image enhancement,image sharpening,edge detection and threshold segmentation.Feature extraction is performed on the image obtained by the pretreatment and the parameters of the software are set in order to judge the quality of the glass bottle according to the extracted image features.The position and range of the glass bottle defect is determined using the image connected region information.In addition,a deep learning algorithm was developed to identify defects such as cracks,bubbles,stones,and cold spots in glass bottles.The main process is to construct a sample glass bottle with different defects into a deep learning data set and construct Le Net-5convolutional neural network and VGG-16 convolutional neural network and evaluate the performance of the network.Train the convolutional neural network using training set data by setting the initial hyperparameters of the network.Analyze the loss function curve of the training set after network training and observe the effect of network training and verify the performance of convolutional neural network using the verification set data.Adjusting the initial hyperparameters by verification results of the neural network.After repeating the training and verification of the convolutional neural network,the performance of the network is adjusted to the optimal state.Finally,using test set data to test the defect recognition accuracy of convolutional neural networks.Comparing the accuracy data of Le Net-5convolutional neural network and VGG-16 convolutional neural network,the recognition performance of the two networks was analyzed.The designed experimental platform and algorithm realize the defect detection of the bottle mouth and bottle body.