| In the industrial production,an automatic dispenser is usually used to dispense the objects that need to be adhered.However due to the influence of many factors,the glue may cause many defects such as break,overflow,lack of glue,etc.In order not to affect the quality of the product,it is necessary to test the quality of the dispensing.Relying on manpower to detect the quality of dispensing has the disadvantages of low efficiency,strong subjectivity and fatigue.In order to solve the above-mentioned defects,this project uses computer vision method to construct an automatic optical detection system for dispensing defects,also known as dispensing defect detection system,to automatically detect whether there are defects and defect categories of the rubber strip.In this paper,according to the requirements of dispensing defect detection system and mechanical structure,the detection process plan,motion control plan and optical detection plan are developed.The dispensing defects are divided into five categories: multi-glue,less-glue,breaking,distortion and bubble.The detection algorithm based on image processing is compared with the deep learning algorithm based on convolutional neural network.In terms of image processing algorithms,for the multi-glue and less-glue defects,the rubber strip image is pre-processed to obtain a binarized image,edge extraction is performed,the minimum enclosing rectangle of the contour is obtained,and the obtained rectangular region is vertically or horizontal adjusted,counting the number of pixel values,judging whether there is a defect according to the difference between the maximum and minimum values of the number of pixel values and the shape of the curve in the projection map.For the breaking glue defect,in order to obtain the contour of the rubber strip at the edge of the image,the image is edge-expanded,the processed binarized image is subjected to a closed operation,contour extraction is performed,and the number of contours and the size of the area are counted as a basis for judging whether or not there is a breaking glue defect.For the bubble defect,since the bubble is characterized by a circular shape,the image of the rubber strip is subjected to a Hough circle transformation,and the number of rounds and the size of the area are statistically detected.The detection algorithm based on image processing can only be targeted to specific application scenarios,lacking versatility,and algorithmic accuracy is not very high.In the aspect of deep learning algorithm,the Lenet-5neural network model is finally used through model comparison.At the sametime,the improved schemes are proposed,which are data enhancement,migration learning and improving network structure.Through experimental data analysis,the model combined with data enhancement and improving network structure has the best effect on the problem of dispensing defect classification,and the accuracy of each type of defect and the test time of single picture are better than image processing algorithm.Therefore,this paper uses the deep learning algorithm based on convolutional neural network as the detection model of the dispensing defect detection system.For the visualization of the model detection process,this paper has written a corresponding software detection page to achieve end-to-end detection. |
