Multiobjective framework for hardware and quality aware approximate Gaussian filtering toward energy efficient...

Area of Science:

• Embedded Systems Engineering
• Medical Imaging
• Computer Vision

Background:

• Real-time ultrasound image denoising is crucial for healthcare but limited by computational costs on embedded platforms.
• Existing methods struggle with the power, energy, and latency constraints of resource-limited devices.

Purpose of the Study:

• To develop a unified, automated design space exploration (DSE) framework for hardware- and energy-efficient ultrasound image denoising.
• To integrate approximate computing with multi-objective evolutionary optimization for balancing image quality, power, area, and delay.

Main Methods:

• A pipelined 2D Gaussian filter architecture using approximate adders/multipliers from the EvoApprox8b library.
• Integration of evolutionary algorithms (NSGA-II, NSGA-III, MOPSO) and random search for DSE.
• Introduction of the Quality-Area-Power Product (QUAP) metric for ranking solutions.

Main Results:

• NSGA-II achieved up to 90.1% reduction in power-area product with high image quality (PSNR > 35 dB, SSIM > 0.93) in three-objective optimization.
• NSGA-III demonstrated superior diversity and robustness under many-objective settings with timing constraints.
• The framework enabled scalable synthesis of approximate Gaussian filtering accelerators.

Conclusions:

• The proposed DSE framework facilitates the creation of energy-efficient embedded imaging systems for ultrasound.
• This approach supports the development of sustainable and accessible healthcare technologies through optimized hardware accelerators.