Research Of Feature Extraction And Selection Algorithm In Capsule Endoscopy Images

Capsule endoscopy (CE) has been widely applied in hospital due to its great advantage on directly obtaining the entire small bowel in patient body compared with traditional endoscopies and other imaging processing techniques for gastrointestinal diseases. However, CE will produce sixteen to eighteen thousand images during each test. Normally, all images should be inspected by physicians with eyes, which cause a great burden to physicians. To solve this problem it is of great importance to develop computer assisted diagnosis system. Feature extraction and feature generation in filtering abnormal images are major issues concerned in this thesis. For each CE image, extracting features with strong differentiation in lesion images has a critical influence on classification performance. A large number of high-dimensional feature vector after feature extraction require minimizing the dimension to ensure the specificity and accuracy of the classification system. However the existing system makes the result unreliable.Feature extraction and feature generation have been mainly studied in this thesis for diseases detection of CE images. The author's main research work and contributions are as follows:Firstly, CE images are color images while different color spaces have different image representations. To better understand the formation of color images and analyze color invariance, the dichromatic reflection model has been used. Color feature is the most convenient image primitive feature and shape feature is the primitive feature which can connect primitive feature with semantic feature. This thesis use nine order color moments and color tensor as images' color feature and Zernike moments to extract images' shape feature.Secondly, CE images have rich texture information and there are significant differences between normal and abnormal images. This thesis analyzes the existing methods including local binary pattern and wavelet transform first, and then uses local ternary pattern which is invariant to noise and Contourlet transform which can capture the directional structure to extract images' texture features.Thirdly, although too many features may carry good classification information when treated separately, there is little gain if they are combined into a feature vector because of a high mutual correlation. Meanwhile, the higher the ratio of the number of training patterns N to the dimension of feature vector, the better the generalization properties of the resulting classifier. This thesis uses feature selection algorithm to generate feature vector with better performance.