Research On Digestive Tract Image Classification Method Based On Active Learning Sampling Strategy

The wireless capsule endoscopic digestive tract image-assisted diagnosis based on deep learning has achieved remarkable results in recent years.During the diagnosis process,it is often necessary to screen and analyze specific digestive tract organs,and the digestive tract data set of a single patient is very large.Therefore,training an effective model for automatic classification of digestive tract organs will greatly reduce the workload of diagnosis and improve diagnosis efficiency.In practical applications,supervised learning technology is easy to achieve higher accuracy,and still occupies a major position.However,supervised learning requires enough labeled data for model training,and it is expensive to collect and label enough digestive tract images.How to reduce the cost of labeling while ensuring the performance of the model is an urgent problem to be solved.This paper first proposes a digestive tract image classification method based on Resnet50.This method uses the fine-tuned Resnet50 model and pre-trained weights on Image Net,and the model is trained without freezing the feature layer.Experiments show that this method has achieved better performance than other classification methods.This method uses 100% of the training set for model training,and achieves 95.41% and 91%classification accuracy on the validation set and test set,respectively.Secondly,the above method is improved,and an improved digestive tract image classification method based on active learning sampling strategy is proposed.This method uses Resnet50 as a classification model and combines entropy-based sampling strategies for active learning.In the sampling process of each round of active learning,the entropy value of the prediction probability of the sample is calculated by the classification model,and it is used as the sample value indicator,and the most valuable batch of samples is selected for labeling and model training.Experiments show that this method can make the model achieve better generalization performance while using less labeled data.This method uses50% of the training set for model training,and achieves 94.58% and 92.28% classification accuracy on the validation set and test set,respectively.Finally,an auxiliary classification system for digestive tract images is designed.The system is presented in the form of a GUI page,and the user can conveniently call the trained algorithm to classify the digestive tract images.The system can be directly used in the diagnosis process to realize automatic classification of the digestive tract organs of patients,and can also be used in the sample labeling stage of active learning to realize pre-labeling of extracted samples.