Research On Optimized Algorithm For Classification Of Visible Particles In Large Infusion Bottles

Large infusion is one of the most commonly used and important pharmaceutical preparations in the pharmaceutical industry.However,due to the manufacturing process and production environment,it is very easy to mix metal filings,mosquitoes,hair,black spots,plastic crumbs and other visible foreign objects in the liquid,which will seriously endanger the lives of users.At present,the country is increasing the prevention and control management of medical infusion drugs,and the requirements for foreign body detection are becoming more and more stringent.Realizing the accurate category of foreign objects can analyze in detail the source,degree of danger and control measures of the foreign objects,and strictly control the source of the foreign objects during the processing process,so as to control the product quality of large infusion bottles from the source of the foreign objects.After actual research,most manufacturers have not realized the detailed classification of foreign bodies,but distinguished into obvious and inconspicuous foreign objects based on the detected size of the foreign objects.So,there is no mature product on the market that classify the visible foreign object of large infusions.Therefore,this paper takes the large infusion packaged in plastic bottle as the research object,expands on the research of foreign object detection,and designs the classification convolutional neural network,and proposes the optimization algorithm to improve the classification accuracy for the difficulty of foreign object analysis.The main work of this paper is as follows:1.There is currently no open-source classification database of foreign objects,so this paper constructed a classification dataset of different category foreign objects which can provide the true and various dataset foundations for relevant classification algorithm by organizing and classifying foreign object images which were obtained by regional crop based on the existing large infusion visible foreign object detection algorithm.2.In view of the many difficulties,such as various types of foreign objects in the foreign object image,the presence of weak and small foreign objects,the difference of characteristics,size,and posture of the same type of foreign objects,and the similar features such as shape,size,and gray scale between categories.This paper proposed using the deep learning convolutional neural network that has strong automatic feature extraction and learning to classify the different types of foreign objects.Then proposed an image classification algorithm based on EfficientNet,and obtained the classification model by training the foreign object dataset,and the classification accuracy of foreign objects is 92.25%.3.In the actual process,the foreign object image data has a serious imbalance between categories,which results in the too low classification accuracy for the minority foreign object and reduces the overall classification performance and greatly affects its use.This paper introduced a cost-sensitive learning method,and uses class balance loss to re-weight the loss function in the training of the classification algorithm to solve the imbalanced classification problem.Comparative experimental results show that this method improves the performance of the foreign object classifier,improves the classification accuracy of minority black spots,paper scraps and plastic scraps in foreign objects data,and increases the overall classification accuracy of the algorithm to 95.50%.4.It can be seen that some categories data of the foreign objects is weak and lack of information such as the shape,size and texture of the foreign body,which leads to the lack of information used by the algorithm.Therefore,this article applied super-resolution technology to the large infusion visible foreign objects classification algorithm,and compared with other algorithms,we found that the super-resolution algorithm based on ESRGAN can generate enhanced foreign objects images with richer details and textures and better human vision by reconstructing the original foreign body image dataset.Then we used the foreign objects image dataset before and after super-resolution enhancement for training and compared the test result of the two obtained classification models.It was verified through experiments that the super-resolution algorithm proposed in paper can promote foreign body classification tasks and effectively improved the accuracy of weak foreign objects classification.Aiming at the difficulty of the algorithm obtained from the analysis of actual foreign objects characteristics,this paper used deep learning as the core technology and focused on improving the accuracy of foreign object classification.The relevant research on the foreign body classification and classification optimization algorithm was carried out,and a series of experiments have verified the effectiveness and practicality of the entire algorithm.The optimization classification algorithm used in this article has greatly improved the classification accuracy and has strong practical and economic value.