Research On Defect Detection Method Of Lotion Pump Based On Machine Vision

The lotion pump is an important part of the lotion container,and its social demand is large,which requires mass production.In the process of production,it will produce defects such as oil stains and scratches,affected by the injection mold and the assembly machine.At present,the industry uses manual detection to detect defects,and manual detection is susceptible to the subjective emotions of the testers,reducing the reliability of the test results.This paper aims to study the defect detection method of emulsion pump based on machine vision,use machine detection instead of human detection,promote the trend of production automation,and accelerate the pace of enterprise transformation and upgrading.The CCD camera is used to convert the detection of the individual pump body into defect detection at the top of pump,the neck of pump,the body of pump,and the tail pipe.On the one hand,qualitative classification and defect detection on the image angle of view are studied.One the other hand,qualitative classification and defect detection on the pump level are studied.The specific research questions are as follows:(1)The model of Qualitative classification of defection.The purpose of this model is to achieve a qualitative two-category of normal and defective images at the perspective image level.Firstly,a qualitative classification model based on traditional machine vision SVM is established.The HOG and LBP features of the sample image are extracted and the above features are input into the training model in the SVM.Then the internal structure of the convolutional neural network DarkNet-53 is studied,and a qualitative classification model based on DarkNet-53 is established.Finally,based on depth-wise separable convolution and group normalization,DarkNet-53 is improved.I design a new residual structure,which makes the model more lightweight,and the batch normalization effect is not restricted by the batch size.The improved model DarkNet has better classification accuracy and detection efficiency than other models,and the average detection time is only 65.23 ms.In addition,the paper also completed the study of image optimal scaling,data augument and visualization of model decision regions based on Grad-CAM.(2)Defect detection model.The purpose of this model is to achieve positional regression and category prediction of defects at the perspective image level.Firstly,the common algorithms for target detection are introduced.Combined with the research task of this paper,the convolutional neural network YOLOv3 is selected as the research object.Secondly,the network model structure of YOLOv3,model output tensor,candidate frame size initialization and prediction frame filtering mechanism are studied,and a defect detection model based on YOLOv3 is established.Then analyze the two shortcomings of YOLOv3: insensitivity to fine scratches and small oil stains and general effects on singlesample and multi-defects.Improvement of YOLOv3 based on qualitative classification model DarkNet: Replace YOLOv3's backbone network with DarkNet,and expand the input size to 736 to improve the model's performance for fine scratches and small oil stain detection.K-means++ is used instead of K-means to complete the clustering task of the label box,and the stability of the clustering result and the average IOU of the candidate box and the truth box are improved.The more robust Soft NMS is used as a new predictive frame screening mechanism to improve the ability of the model to detect single-sample and multiple defects.Finally,the weight migration of the qualitative classification model is studied.The weights of the trained qualitative classification model are used for the weight transfer of the defect detection model to improve the training efficiency of the model.The training strategy based on loss callback is studied,the network training is stopped early,and the parameter over-fitting is reduced.(3)Lotion pump qualitative classification and defect detection.Achieve qualitative classification and defect detection at the pump level.Firstly,a human-computer interaction platform is built to control the model detection process to achieve qualitative classification and defect detection at the pump level.Then the qualitative classification effect of the test method on the lotion pump and the detection performance of each type of defect.The accuracy of qualitative classification is 98.4%,and the average detection time of each pump is 0.726 s.It can completely detect defections such as hungry joint and reverse insertion,and the recall rate of scratches is lower than other defections,but still can reach 92.65%.It can meet industrial indicators that test at least 60 pump bodies per minute,the qualitative classification accuracy is not less than 95%,and the recall rate of each type of defects is higher than 90%.Finally,consider the industrial application of the method,and analyze the main flow of the method application.