A Quantized Convolutional Neural Network Implemented With Memristor for Image Denoising and Recognition.

Area of Science:

• Materials Science
• Computer Science
• Electrical Engineering

Background:

• Image noise degrades quality, impacting processing and visual appeal.
• Current denoising algorithms are computationally intensive and energy-consuming.
• Memristor devices offer potential for efficient hardware solutions.

Purpose of the Study:

• To propose an efficient image denoising and recognition method using memristor multi-conductance states.
• To demonstrate the feasibility of memristor-based convolutional neural networks (CNNs) for image processing.
• To guide the development of high-performance image noise reduction hardware.

Main Methods:

• Utilized Pt/ZnO/Pt memristor wires to achieve 26 continuous conductance states.
• Mapped memristor conductance states to image pixels for feature preservation and noise reduction.
• Implemented weight quantization of CNNs based on multi-conductance states.

Main Results:

• Successfully mapped conductance states to image pixels, enabling noise reduction and feature preservation.
• Demonstrated the feasibility of using multi-conductance states for CNN-based image denoising and recognition.
• Achieved efficient image processing through memristor device integration.

Conclusions:

• The proposed memristor-based method effectively reduces image noise while preserving features.
• Multi-conductance states of memristors are suitable for CNN weight quantization.
• This approach offers a promising direction for energy-efficient, high-performance image noise reduction hardware.