A Study Of Image Similarity Based On Multi-features

The similarity of the images based on the research about the basic of properties of the image is studied in this paper. The image similarity is the key of the image retrieval and the image classification. At first the paper gives the definition of similar images and introduces three categories of similar images. Then study the image similarity from color,texture and shape.The color feature vectors are extracted by splitting the image to rings and then doing histogram statistics in HSV mode. The feature vector includes mean,standard deviation,skewness,energy and information entropy. The HSV mode which defines the color from the perspective of human vision is more suitable to extract the color feature. The operation of splitting the image to rings not only makes up for the shortcomings of the histogram statistics which is lack of spatial information, but also maintains the rotation invariance of the feature vectors.The texture feature vectors are extracted from the gradient image by co-occurrence matrix statistics. The feature vector includes energy,contrast,information entropy,inverse gap and correlation. Because that the co-occurrence matrix is lack of spatial information, we do co-occurrence matrix statistics from four directions from four gradient images.The shape feature vectors are extracted from the image contour by co-occurrence matrix statistics. The shape of image has two representations:contours and regions. Because of having strict spatial information, the regional characteristic is not suitable for statistics, so the contours is used to do co-occurrence matrix statistics. The feature that the contour itself has spatial information not only makes up for the shortcomings of the co-occurrence matrix statistics, but also enhances the spatial information of the co-occurrence matrix statistics.The color feature vectors are useless in processing some similar images.The texture and shape feature vectors has miscarriage of justice areas. So an integrated algorithm combined with the advantages of each aspect has been created to identify similar images better and more accurate. This algorithm first filter out the images which is suitable for method of the color feature vectors. And then process the other images by the texture and shape feature vectors. The flag calculated from the miscarriage of justice areas of the texture and shape feature vectors is used to detect the image similarity. This algorithm has good adaptability, high accuracy and high robustness.