Research On Pill Classification And Recognition Based On Deep Learning

In recent years,people’s living standards have gradually improved,and the pursuit of pharmaceutical quality has become increasingly stringent.The special nature of pharmaceuticals makes it necessary that they should be supervised and inspected at all stages from research and development,production,packaging to distribution,etc.The quality classification and identification of tablets are particularly important.How to efficiently sort out and reject unqualified drugs at the time of drug production and delivery and before packaging and storage is an important direction of research in the field of drug production.The traditional method of manual sampling is inefficient,error-prone,and expensive,and cannot meet the development requirements of industrial automation.Aiming at the phenomenon of unqualified tablets due to imperfect detection during the production of tablets in existing enterprises,this dissertation proposes a tablet classification and identification method based on an improved EfficientNet-B0 network,which can efficiently identify unqualified tablets before they are produced and packaged.The main work is as follows.1.Investigate the EfficientNet network based on it.Analyze the scaling idea,internal module MBConv composition architecture,and channel attention mechanism idea of the EfficientNet network.Explore the internal composition structure of the network as well as the performance advantages and disadvantages and provide the theoretical basis for the classification recognition research later.2.For the classification recognition of pills,the baseline network EfficientNet-B0 is improved to obtain higher recognition accuracy and computation rate.The improvement measures include: simplifying the network structure module,retaining the core architecture of the network,reducing the network parameters,simplifying the computation,achieving a good classification recognition effect;introducing the ECA module,improving the SE module into the ECA module,avoiding the side effects of dimensionality reduction,and adaptively acquiring the weight information between channels through one-dimensional convolution to improve the accuracy and operation rate of the network model.3.The pill image dataset is constructed,the dataset is expanded by image enhancement operation,and the image preprocessing performs size normalization and other operations on the dataset.The dataset is randomly divided into a training set,validation set,and test set in the ratio of 70%:15%:15%.4.The improved EfficientNet-B0 network is used to test and evaluate the performance of the pill dataset,specifically including the performance evaluation of the network structure,the performance comparison of Google Net,Moblie Net V2,Conv Ne Xt,and the original EfficientNet-B0 algorithm,and the comparison of the classification accuracy of different classes of pills.The experimental results show that the improved model achieves98.93% pill classification accuracy,which is 1.83% higher than the original network,shortens recognition time,improves real-time efficiency,and achieves better classification results.It provides a reference solution for the convolutional neural network in pill classification recognition.