Area-efficient low-power image watermarking architecture using faithful approximation and reversible logic
Silambarasan D.,Vijeyakumar K.N.,Saranya K
DOI: https://doi.org/10.1007/s11042-024-19702-x
IF: 2.577
2024-07-18
Multimedia Tools and Applications
Abstract:Approximate computing and reversible logic have become vital concepts in designing low-power, area-efficient datapath units for portable image and signal processing applications. In this approach, a low-power, reversible architecture targeting digital image watermarking is proposed. The reported low-power Reversible Watermark Architecture (LP-RWA) implements a novel high-speed embed unit to blend the watermark image onto the cover image. The proposed embed unit performs block-level parallel accumulation and incorporates hybrid computing with approximate parallel carry logic (APC) for n/2 least significant bits (LSBs) and the standard carry logic (SC) for n/2 most significant bits (MSBs) to achieve area efficiency and speed improvement. Note that the implementation of APC in the least n/2 bits introduces a maximal error of 2 n/2 ; however, we add a suitable constant unit bias at an appropriate position to faithfully restrict the error within -2 n/2–1 to + 2 n/2–1 . Furthermore, we use reversible logic for implementation to achieve low-power operation and perfect recovery of the watermark image from the blended image without much signal loss. An area-efficient version of the LP-RWA (AE-RWA) is also reported, eliminating the generation of least significant n/4 bits in the approximate part of the Reversible Faithful Embed (RFE) unit. This elimination operation increases the average error during recovery; however, the maximum error is maintained within ± 2 n/2–1 . Evaluations show that LP-RWA exhibits 38.8%, 25%, 27.7%, and 27.2% reductions, and AE-RWA exhibits 46%, 32.7%, 34.3%, and 58.6% reductions in Quantum Cost (QC), Gate Count (GC), Garbage Output (GO), and Logic Depth, respectively, when compared to the standard reversible watermarking architecture. Synthesis results of LP-RWA and AE-RWA designs with 90 nm Application Specific Integrated Circuit (ASIC) Process Design Kit (PDK) technology with a number of bits in input operand (n) = 16 and employing 4 bits in each block showed 70.2% and 72.8% power delay product (PDP) reductions, respectively, when compared to the standard design. Evaluations through FPGA implementations and standard test images reveal that the LP-RWA and AE-RWA generate processed outputs with SSIM values of 0.9793 and 0.9545 and peak signal-to-noise ratio (PSNR) values of 37.77 dB and 36.2, respectively, on an average. Furthermore, analysis of LP-RWA and AE-RWA with different malicious benchmark attacks proved their high robustness.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering