Design Of Information-Theoretic Image Similarity Measures

Image similarity or distortion assessment is fundamental to a broad range of applications throughout the field of image processing and machine vision. Many existing image similarity measures have been proposed to handle specific types of image distortions. Also, there are methods such as the classical structural similarity (SSIM) index that are applicable to a wider range of applications. Most of existing image similarity measures are based on statistical approaches. Image similarity measures that are based on information theory are getting attention because of their capability to predict relationships between image intensity values.In this thesis, we propose, analyze, and test two novel techniques for image similarity using information-theoretic approaches. Both techniques work well to detect similarity under noisy conditions. The first measure (named as HSSIM) is proposed for general similarity testing between various kinds of images. The second measure (named as ISSIM) is proposed for a specific application in face recognition. The new similarity measures are based on joint histogram.SSIM can produce confusing results in some cases where it may give a non-trivial amount of similarity while the two images are quite different. Also, it fails to detect similarity at low peak signal to noise ratio (PSNR).Therefore, we are motivated to propose the first similarity measure HSSIM, by using information-theoretic technique based on joint histogram 2D self-symmetry, normalized with respect to the test image 1D histogram. The proposed method has been tested under Gaussian noise. Simulation results show that the proposed measure HSSIM outperforms statistical similarity SSIM by ability to detect similarity under very low PSNR. The average difference between SSIM and HSSIM can reach about 20 dB.ISSIM is a novel technique for face recognition. This technique is an extension to our information-theoretic technique utilized in our first similarity measure HSSIM. Face recognition with ISSIM is based on classifying intensity values of gray scale pixels into different groups. Face recognition is a classical unresolved problem in image processing, well-known to be time-consuming. However, our information-theoretic approach is efficient as it has limited computations as compared with other approaches, making it suitable for image (face) recognition in real-time environment. For performance evaluation of ISSIM, we proposed a new performance measure that can reveal the strength of recognition. The new measure is based on the difference between the similarity of the test image with the recognized face and its similarity with the second-likely face in the database.Evaluation was performed using MATLAB. For HSSIM, we used three kinds of images:landscape, face, and specific shapes (like coins). Gaussian noise was used to test strength of the approach as compared to the classical SSIM. For face recognition using ISSIM, we used a subset of the standard AT&T face-image database comprising 49 face images. The subset contains seven subjects and each subject having seven views (poses) with different facial expressions.Performance of our proposed methods HSSIM and ISSIM is demonstrated experimentally and is shown to outperform the statistical method (SSIM).