BMR and BMW Metaphor-less Optimization for Robust Medical Image Watermarking in Telemedicine

Abstract

Medical images are frequently stored and exchanged through digital systems, making copyright protection and watermark recovery important requirements in telemedicine workflows. This study presents a blind, content-adaptive medical image watermarking framework in the 8×8 discrete cosine transform (DCT) domain. A 32×32 binary copyright logo is embedded repeatedly in the DCT blocks by exchanging two low-frequency AC coefficients. Instead of using one fixed embedding strength, six host-activity classes are formed from block variance and a separate strength is optimized for each class. The optimization is evaluated with five metaphor-less population methods: Rao-1, Rao-2, Rao-3, Best-Mean-Random (BMR), and Best-Worst-Random (BWR). The fitness function combines host-image PSNR with normalized correlation (NCC) obtained after Gaussian noise and JPEG quality-50 attacks. Experiments use a 512×512 grayscale medical cover image and a 32×32 binary watermark, with fourfold redundancy during blind extraction. The supplied implementation reports PSNR values of 38.65, 38.59, 38.59, 38.53 and 38.40 dB for Rao-1, Rao-2, Rao-3, BMR and BWR, respectively, while SSIM values range from 0.9478 to 0.9521. Across the tested attacks, NCC remains around 0.965 for most non-geometric operations, whereas the 2° rotation attack reduces NCC to 0.2703–0.3002. The results show the strength of the proposed blind DCT design against several common signal-processing operations, while also exposing the need for explicit geometric synchronization in future work.