Hardware-aware implementation of FLBMF-selective mean-TV for mammogram denoising on low-power edge devices: FPGA validation and energy efficiency analysis
Benard Nyangena Kiage, Michael Waema Kimwele, Wilson Kipchirchir Cheruiyot · Journal of Intelligent & Fuzzy Systems · 2026
Building on prior algorithmic work that introduced the FLBMF-selective Mean-TV hybrid framework for mammogram denoising, this paper presents the first hardware-aware implementation of an enhanced FLBMF-Selective Mean-TV architecture specifically designed for low-power edge devices. The current manuscript contributes three major advances beyond previous work: (1) replacement of median filtering with cached mean filtering that reuses pre-computed fuzzy similarity weights from the detection stage, reducing per-pixel operations by 33%; (2) a streaming row-buffered architecture that reduces on-chip memory footprint from >256KB to 3.8KB for 512 × 512 images; and (3) the first FPGA implementation of this framework on Xilinx Artix-7, achieving real-time throughput of 3.2 frames per second at 2048 × 2048 resolution with measured power consumption of 0.7 W. Experimental validation using mammogram images from the DDSM database, corrupted with impulse noise ranging from 40% - 90%, confirms that the proposed implementation maintains denoising performance (PSNR up to 32.4 dB at 40% noise, 23.0 dB at 90% noise; SSIM up to 0.96; FOM up to 0.98) while operating within strict resource constraints. The design requires only 12 integer operations per pixel on average, achieving 79% computational savings compared to Non-Local Means filtering. All arithmetic operations use integer-only fixed-point approximations with bit-shift substitutions for division, eliminating the need for floating-point hardware. These results demonstrate that the FLBMF-Selective Mean-TV framework can be deployed effectively on resource-constrained edge platforms, enabling real-time and low-power mammogram denoising for intelligent healthcare applications in portable and remote screening environments.