Study On Segmentation And Recognition Methods Of Anal-Intestinal Diseases Images

In medicine, hemorrhoids are common and vulnerable diseases. There are manyChinese suffering from the diseases. The traditional diagnostic methods ofhemorrhoids are based on those doctors’ own expertise and clinical experience formany years to diagnose diseases. This method requires the doctors’ expertise in adown-to-earth manner, rich clinical experience, and is easy misdiagnosis, lowproductivity[1]. With the continuous development of computer science and artificialintelligence subject, in order to improve the accuracy and their intelligent level ofanal-intestinal diseases recognition, many medical experts in the field hope that theanal-intestinal diseases can be automatically identified by using expert system. On theabove considerations, this paper mainly study on segmentation and recognitionmethods of anal-intestinal diseases images.Firstly, select the images which are taken in clinical medicine, and the imagesbrightness is adjusted by Grey-World, Shades of Grey, max-RGB, and Grey-Edgealgorithms. The adjusted images are converted into gray ones and then dealt with bymedian filter and mean filter for denosing.Secondly, it is necessary to segment images in order to study the interested areasand eliminate some irrelevant factors. On the basis of image preprocessing, imagesare segmented by semi-threshold segmentation, improved one-dimensional maximumentropy thresholding method and two-dimension maximum entropy thresholdingmethod. Through many experiments, this paper finds that semi-thresholdsegmentation method is efficient to segment the images. And it is very effective forextracting texture feature by the segmented images with this algorithm. In order toextract the moment invariant features of images, segmented images edges aredetected by using krisch algorithm.Then, extract the segmented image features. In this paper images’ statisticalcharacteristics、shape and texture characteristics are extracted. In the statisticalcharacteristics, image gray mean value, variance, entropy are selected. In terms ofshape features, Hu invariant moments are selected. In terms of texture features gray-level co-occurrence matrix and gray-gradient co-occurrence matrix are selected.The dimensions of extracted feature data are reduced by PCA(Principal ComponentAnalysis)method, and the extracted main components are put into the classifier fortraining. According to the recognition methods which are provided in patternrecognition, the paper selects SVM (Support Vector Machines, SVM) and BP neuralnetwork to train data and recognize targets.Finally, the experiment simulations are done by using two identificationmethods on Matlab platform. Based on the simulation results, the rate of BP NeuralNetwork recognition algorithm is slightly higher than the SVM algorithm, butconsidering the small sample size in the subject, stability of recognition method, therecognition method based on SVM is more suitable for the identification of thesubject.