Research And Application Of Medical Sample Container Recognition Based On Convolutional Nerual Network

In automated diagnostic equipment,the accurate recognition of samples is the basis of detection,the recognition of sample is the recognition of the sample container in which samples are stored.At present,almost all automated diagnostic equipment in China,the recognition of sample containers requires manual intervention by clinical laboratory technologist.Manual operation will bring risks,giving wrong reports and operator infection.The image recognition technology applied in existing automated diagnostic equipment is using a camera and convolution neural network,which recognition accuracy rate is above 99.5%,but its high price and high demand for hardware resources make it urgent to find an alternative.Therefore,it is necessary to develop a set of “small,fast,accurate” convolutional neural network to achieve the goal of reducing costs and improving integration with equipment.Through the analysis of working principle of convolutional neural network,this thesis absorbs the idea of separable convolution in Mobile Net,which greatly reduces the parameters and computation.Instead of pooling,a convolution core with a stride of2 is used to reduce the structural scale and parameters of the network.The use of global average pooling effectively solves the problem of parameter explosion caused by fully connected layer.Based on the above three points,the scale of parameters and calculation is effectively controlled.The application implemented in this thesis needs to recognize different types of sample container,use the same image data set as the existing solution uses to train,validate,and test.Because the amount of image data in different categories in the data set is quite different,first use a sample balance strategy to process these types of data,so that each type of image data has the same amount of data.Second,the processed image data is disordered to ensure that each batch of image data contains these types of data and the amount of image data is basically equal in the following training,validation,and testing process.Then,the disordered image data is segmented into training set and testing set according the proportion of 8 to 2.The training set uses the cross-validation method to split into 4 training sets and one validation set.Third,data augmentation technology is applied to improve the generalization ability of the network.Finally,the image data is normalized to make all images are within a scale,avoiding singular data and accelerating the convergence of the network.In the training process,by observing the change trend of the loss value of the training set objective function and the verification set objective function,analyze the potential reasons,gradually add improvement measures and combine network training skills,it is finally confirmed that the fourth network is the final network.The final improved convolutional neural network only takes 216 KB in size,and the average recognition time for single image data is 36 milliseconds,and the recognition rate exceeds 99.5%,achieving the goal of developing a "small,fast and accurate" convolutional neural network.