Research On Application Of Image Segmentation,Feature Clustering And Feature-class Matching

Image segmentation,feature clustering and feature-class matching are important research topics in the fields of image processing and visual computing.We discuss the methods of image segmentation,feature clustering and feature-class matching and apply them to facial images and diabetic retinopathy images in this paper.We first discuss the methods of feature clustering and feature-class matching,design an improved algorithm for calculating region similarity,and design a feature-class based matching method for facial images.The specific contents include:(1)The image segmentation method based on Harr-like features is used to extract the face region in input images to eliminate the influence of backgrounds on the matching results.(2)We carry out threshold selection experiments on relevant parameters in SIFT algorithm to ensure that we get as many feature vectors as possible without affecting the matching accuracy.Thus,we can get more stable facial features.(3)The SIFT feature extraction algorithm is applied to face images,and the distribution of SIFT features on face images is observed.Since the features are mainly distributed in five areas,i.e.,the two eyes,the nose and the two comers of mouth,we select five initial centers to cluster SIFT features by K-means algorithm.Thus,we complete the area division of face images and match features between the same areas of different images.(4)We design an improved similarity measure formula by combining the Euclidean distance and the angle cosine based on the common similarity measure formula in vector space.(5)We assign different weights to different areas in face images in consideration of each area having different influence on matching results,and give a feature matching method for face images based on feature-class matching.We carry out feature-class based face matching experiments on the publicly available face image databases,ORL and YALE.The experimental results illustrate the effectiveness of our method.We then discuss the methods of locating optic disc in diabetic retinopathy images to improve the computer-aided diagnosis efficiency of retinopathy in fundus images in this paper.The specific contents include:(1)We use the threshold segmentation method to do two-value processing of fundus image,and extract ROI of fundus images to reduce the amount of computation and eliminate the influence of black backgrounds on the locating results.(2)By trying a variety of luminance normalization schemes,we select the normalization method based on proportional operators to reduce the influence of uneven brightness and contrast in fundus images on the locating results.(3)Through analyzing the grayscale processing results of fundus images,we find that the green channel can achieve high contrast while retaining more complete image information.Based on this,we design a general optic disc template for diabetic retinopathy images.(4)We give a fast optic disc locating method based on the optic disc template for diabetic retinopathy images.In order to locate the optic disc in a fundus image,we calculate the similarity measurements between the optic disc template and each coverage area in the fundus image by sliding the template so as to determine the most similar area.We carry out matching and locating experiments on the largest fundus image database MESSIDOR.Extensive experiments show that the template based fast matching method can effectively locate the optic disc in fundus images.Our work can improve the performance of image matching to some extent and has significance in theory and practice.