Study Of Classification Algorithms For Capsule Endoscopy Small Intestine Images

As a novel technology, Wireless Capsule Endoscopy (WCE) has been widely used in human intestinal disease diagnosis. Although it is continuously improved with the technological advancements, there are still difficulties in practical applications. Huge number of recovered endoscope images must be diagnosed by professionals, it is a very time consuming process. As the results, the automatic computer aided disease diagnostic system is highly required.The diagnostic classification algorithms are studied in WCE small intestinal images, and they are the major issues concerned in this thesis. This thesis analyzes the existing differences between normal and abnormal WCE images, and extracts color and texture features from these images. And then machine learning methods are used to learn and classify the features in the thesis. Finally, the purpose of the thesis is to find out a suitable and accurate classification algorithm. The author’s main research works are as follows:Firstly, this thesis analyzes color and texture features of every image in the different color spaces. This thesis uses the color moment as images’color feature and compares the different order of the color moment. This thesis also analyzes the existing texture information methods, and it contains local binary pattern and Contourlet transform. Using the features fusion method to analyze the performance. One is the color moment and Contourlet transform, the other is the local binary pattern and Contourlet transform.Secondly, Gaussian process classification algorithms are implemented in this thesis. Based on the choice of Gaussian process model, this thesis studies two Gaussian process classification approximation algorithms. Different kinds of Gaussian process classification methods compared with the traditional supervised learning. The experimental results show that the proposed method achieves better performance in WCE small intestinal image classification.Thirdly, Multi-Instance Learning is introduced to the disease diagnostic system. Jn order to improve the traditional diagnostic system, which is extracted the global features. A cascade diagnostic classification method is proposed in the thesis, which is extracted the global features and the locality features. So that it can release the film reading doctors’ workload and improve the diagnostic efficiency.