Capsule Endoscopy Image Recognition Based On Fusion Of Deep Feature And Traditional Feature

Gastrointestinal related diseases are becoming a huge threat to our health,chose the early endoscopic diagnosis and treatment techniques,diagnosis and treatment in earlier stage,it can Significant reduce the probability of deterioration.The endoscope is an essential equipment for doctors to examine patients.It is ingeniously designed,and it is constantly growing and updating.One of the more promising developments is the capsule endoscope(CE).A large amount of endoscopic video or image data will be generated during CE examination.If it is only manually marked or identified,it will take a lot of time and manpower.In addition,the structure of human gastrointestinal organs is very complicated,and the abnormal type pictures look different.The diagnosis is very difficult.The purpose of this paper is to propose an endoscopic image detection classification algorithm,which can realize automatic and accurate classification of endoscopes to reduce the workload of medical staff and speed up the diagnosis.The classification methods of previous studies are mostly based on the traditional characteristics of people’s experience and cognitive design.These characteristics are not universal,the types of lesions that can be classified are limited,and the classification accuracy needs to be improved.Aiming at the multi-lesion CE dataset in the paper,the paper first proposes a classification method that combines many traditional features,selects different types of features and uses multiple feature fusion strategies,in order to fuse more useful information,and finally draws based on The multi-feature fusion classification method can better adapt to the multi-classification task of capsule endoscopic images.In recent years,artificial neural networks,especially deep convolutional neural networks(CNN),have become a research hotspot because they can extract high-level semantic information,and their adaptability and versatility are better than traditional features.However,deep learning requires sample data.Higher,it is easy to over-fit in some data sets(especially medical data sets),and the top-level features extracted by the deep learning model are weaker for local characterization and geometric invariance than traditional features.Another study in this paper is to choose the appropriate depth feature and traditional feature fusion method,so that the final model can combine the advantages of the two feature models and improve the adaptability of the feature to the data set to meet the expected requirements of the article.The research shows that the feature fusion method can better achieve feature complementation in the classification task of multi-class capsule endoscopy dataset.The method of using depth feature and traditional feature fusion is better than CNN feature and traditional feature classification.The method,even the combination of traditional features,finally increases the small-scale running time and significantly improves the capsule endoscope image recognition accuracy index,which shows the broad application prospects of this paper.