Application Research Of Medical Waste Classification Based On Deep Learning

The outbreak of new coronavirus in late 2019 has lasted for two years.It is still the biggest problem in the global medical industry and occupies a lot of resources.In the background of fighting against the epidemic,medical staff bear higher health risks at work than usual,and accidents caused by new coronavirus and other infectious diseases occur frequently.In order to reduce the infection caused by medical waste,this thesis uses the object detection technology based on deep learning to design a medical waste classification system,which can reduce resource waste and labor costs,improve processing efficiency,and reduce the risk of staff infection.The main work of this thesis is as follows :(1)Establish medical waste dataset.Firstly,ten kinds of common medical waste were selected according to the Medical Waste Classification List,and a sufficient number of pictures were obtained through network collection and manual selection.Then,the pictures were annotated and the VOC format medical waste data set was established.Then,a variety of data enhancement methods are used to expand the data set.Finally,the medical waste data set built in this thesis has a total of 10 categories and 12099 pictures.(2)Designed a improved YoloV4 network.Due to the variety of medical waste,in order to improve the accuracy of object detection model,this thesis gives an improved YoloV4 object detection network.The network mainly introduces the Hard-Mish activation function in CSPDarknet53,and introduces the CBAM attention mechanism in the SPP and PANet parts of the main feature extraction network.Similar to Hard-Swish,the hard-coded version of the Mish activation function Hard-Mish is used to replace the activation function in the network.Use spatial attention and channel attention to compose CBAM＿Block.The object detection model with better detection effect is obtained.(3)Training network model and realizing embedded application.Firstly,the K-means method is used to cluster the data set by prior frame,and then the training set and test set are randomly divided according to the ratio of 7.5 : 2.5.In the process of model training,chord annealing method is used to adjust the learning rate and label smoothing method is used to improve the training effect.Finally,the mAP 85.71 % model is obtained.The GUI interface is designed and implemented,which can be used for object detection in multiple ways,such as single detection,video detection and camera detection.Finally,it is deployed on raspberry pi4 B,and the embedded application is realized through remote control of raspberry pi by computer.