Research On Blister Capsule Packaging Defect Detection Technology Based On Machine Vision

The supervisory departments and the market are raising their standards for pharmaceutical industry quality along with the industry’s market’s ongoing expansion.According to Article 213 of the Standards for Quality Control of Pharmaceutical Production,intermediate control checks of pharmaceuticals during the period of pharmaceutical packaging must include ensuring that the pharmaceuticals and packaging materials are correct,the packaging is intact,and the functions of online monitoring devices are normal.The quality of pharmaceutical products is being demanded by the market more and more.Market demands for the pharmaceutical industry’s products are getting higher and higher.Because it relies on high-resolution cameras to capture product images and high-performance industrial processors to interpret those images fast,machine vision detection systems are frequently employed in the pharmaceutical sector to detect the integrity of product packaging.In this paper,a computer software for defect detection of aluminum plastic foam cap packaging based on Goog Le Net network model is developed by analyzing the research status of defect detection of pharmaceutical packaging at home and abroad,comparing the existing research methods,and optimizing and improving the deep learning algorithm.To locate the drug plate to be detected,the improved gray value projection method is first used,using the batch number region of the drug plate as a template.Next,the capsule blister area of the drug plate is divided using this method,and the image of the aluminum-plastic blister capsule area is segmented,and finally,the improved Goog Le Net network model is trained and tested,allowing for the realization of defects such as missing capsules,concave caps,and double caps on blister capsules made of aluminum and plastic.The primary research materials comprise:(1)The camera and light source are chosen,the hardware platform for acquiring images of medicine plates is setup,and the development platform and technical path of the higher computer software are determined in accordance with the visual inspection task of the aluminum plastic blister capsule medicine plate.(2)The pharmaceutical plate image was collected,the batch number area was used as a template,the normalized product correlation gray matching method was combined with the pyramid search strategy to locate the drug plate to be detected,the ROI region of the pharmaceutical plate was used to calculate-clip optimization to narrow the search area,Then,the capsule cap area of the pharmaceutical was separated by the improved gray projection method.(3)The multi-layer perceptron network,a traditional machine learning technique,is utilized to detect it for the data set in this research after feature extraction.The deep learning approach is utilized to compare the detection effects of various network models in the data set in this paper,and a network model improvement optimization based on Goog Le Net is proposed because traditional machine learning method relies on feature engineering and the network generalization is weak.To increase the recognition accuracy of challenging to learn categories like capsule concave caps,the network uses the fusion of attention mechanism to concentrate on the capsule area in the image and add category weights to the cross-entropy loss function.The network model comparison experiment and the ablation experiment demonstrate that the improved and optimized network outperforms the conventional multilayer perceptron and deep learning network models in terms of detection performance.(4)Develop computer software using Winform and Halcon.Configure digital camera on the computer to capture images of the aluminum plastic capsule blister pharmaceutical plate.Set a batch number area template for the pharmaceutical plate and locate the tested capsule blister pharmaceutical plate using the batch number template.Use the computer software to deploy the trained network model and predict the segmented image of the capsule blister area.According to the experimental findings,the improved gray scale projection algorithm obtains a segmentation accuracy of 100 percent for the capsule blister area,and the improved network model has a defect recognition accuracy of over 99%.The improved gray value projection algorithm has high robustness and good segmentation effect.The improved deep learning network model has significantly increased the accuracy of identifying defects in the packaging of aluminum plastic blister capsules compared to the previous one and can be applied to the quality inspection of aluminum plastic blister capsule packaging.