An Improved Reversible Image Transformation Using K-Means Clustering and Block Patching

Recently, reversible image transformation (RIT) technology has attracted considerable attention because it is able not only to generate stego-images that look similar to target images of the same size, but also to recover the secret image losslessly. Therefore, it is very useful in image privacy protection and reversible data hiding in encrypted images. However, the amount of accessorial information, for recording the transformation parameters, is very large in the traditional RIT method, which results in an abrupt degradation of the stego-image quality. In this paper, an improved RIT method for reducing the auxiliary information is proposed. Firstly, we divide secret and target images into non-overlapping blocks, and classify these blocks into K classes by using the K-means clustering method. Secondly, we match blocks in the last (K-T)-classes using the traditional RIT method for a threshold T, in which the secret and target blocks are paired with the same compound index. Thirdly, the accessorial information (AI) produced by the matching can be represented as a secret segment, and the secret segment can be hided by patching blocks in the first T-classes. Experimental results show that the proposed strategy can reduce the AI and improve the stego-image quality effectively.