Method for automatically capturing tissue or foreign matter feature pictures in digestive tract in batches

Abstract

The invention discloses a method for automatically capturing tissue feature pictures in a digestive tract in batches. The method comprises: performing format conversion on a video, removing the background in the video frame, performing gray scale conversion and binarization processing on the target feature, using contour detection of the target feature, outputting the intercepted image of the target feature, storing the intercepted image. The method has the beneficial effects of rapidness, accuracy and convenience.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a method for automatically batch-capturing feature pictures of tissues in a digestive tract. Background technique [0002] Intelligent assisted diagnosis and treatment under digestive endoscopy adopts deep learning as the most effective algorithm to achieve intelligence, and deep learning relies on feature data sets and trained models. In most cases, deep learning models cannot be completely derived from arbitrary data. Learning requires labeling and classification of data. Usually, the labeling and classification of data is performed by personnel who are proficient in target features. However, manual capture and screening of target pictures in videos requires a lot of manpower, but manual The accuracy of the intercepted pictures is low. For pictures with the same characteristics, if the intercepted area, size and fragments are different, it will affect the model traini...

AI Technical Summary

Problems solved by technology

[0002] Intelligent assisted diagnosis and treatment under digestive endoscopy adopts deep learning as the most effective algorithm to achieve intelligence, and deep learning relies on feature data sets and trained models. In most cases, deep learning models cannot be completely derived from arbitrary data. Learning requires labeling and classification of data. Usually, the labeling and classification of data is performed by personnel who are proficient in target features. However, manual capture and screening of target pictures in videos requires a lot of manpower, but manual The accuracy of the intercepted pictures is low. For pictures with the same characteristics, if the intercepted area, size and fragments are different, it will affect the model training of machine learning, and the environment in the digestive tract is non-geometric, dynamic, and has fractal structures. The space of the closed pipeline, the digestive endoscope moves in it, and the identification target tissue is usually placed on the intestinal tract, which causes the input training data of the identification tissue feature to be polluted by the characteristics of the inner wall of the digestive tract, resulting in overfitting phenomenon in the prediction process

Images