Research on cervical cell image classification based on knowledge distillation
Cervical cancer screening is of great significance for cervical cancer prevention and early diagnosis.A knowledge distillation based cervical cell image classification method is proposed to address the issues of insufficient generalization ability,large parameter quantity,high hardware requirements,and difficulty in deploying terminals in existing cervical cell image classifi-cation models.Using residual networks as the backbone network,ResNet18 as the foundation student network,and introducing knowledge distillation mechanism,ResNet34 is used as the teacher network for guiding learning.Adopting transfer learning meth-ods to improve the benchmark accuracy of teacher models;Transfer teacher network probability prediction knowledge through knowledge distillation to student networks for learning,in order to improve the accuracy of student model classification;The experi-mental results show that the accuracy of the student network ResNet18 optimized by knowledge distillation is as high as 95.59%,which is 4.46 percentage higher than the accuracy of 91.13%before optimization.The model optimized by distillation has a small number of parameters,high accuracy,and excellent overall performance of the network,providing a reference for the establishment of clinical lightweight cervical cell image classification models.
NETL
NSTL
万方数据
