Perceptual Similarity Metric Resilient to Rotation for Application in Robust Image Hashing

We propose an objective measure to evaluate perceptual similarity between an original image and its modified version. An input image is first decomposed into a collection of diametric strips to form a secondary image. The secondary image patterns, one of which is shifted and wrapped column-wise for alignment, are low-pass filtered and divided into overlapping blocks. Similarity between the corresponding blocks is then found by calculating correlation coefficients, which are then mapped to the interval. Based on the re-mapped correlation coefficients, the final similarity metric is obtained by dividing a product of several smallest coefficients by that of the same number of largest coefficients. Experimental results show that the proposed metric is insensitive to content-preserving processing such as JPEG compression, rotation, low-pass filtering, watermarking embedding, and moderate noise contamination. On the other hand, it is very sensitive to malicious modification that may only affect a small area of the image. The proposed metric is useful in applications such as image hashing and content-based image retrieval (CBIR).