A Joint Histogram - 2D Correlation Measure for Incomplete Image Similarity

摘要

A hybrid similarity measure is proposed for evaluate the similarity between gray images. The well-known Structural Similarity Index Measure has been designed using a statistical method that fails under high noise. The proposed scale, referred by JhCorr2, uses a mixture of two parts: the first part is based on a two-dimension correlation, while the second part is information - theoretic that using concept of joint histogram with original histogram. The new measure shows the features of information - theoretic approaches and statistical approaches. The proposed similarity measure is robust under conditions of noise and incomplete information. The new measure superior the classical structure similarity and correlation in detecting image similarity at low PSNR under Gaussian, Impulse and multiplicative noise. A face image database AT&T (The image size is 92×112 pixels, with 256 grey levels per pixel) and the well-known image similarity techniques SSIM, corr2 and Jhcorr2 are considered. Peak signal to noise ratio (PSNR), window length in pixel and maximum burst length in pixel were used in this test.