Memristor-Based Selective Convolutional Circuit for High-Density Salt-and-Pepper Noise Removal

TL;DR

This work addresses denoising of images corrupted by salt-and-pepper noise using a memristor-based selective convolution (MSC) circuit that implements SeConvNet-like processing in analog hardware. The authors map ternary SeConv weights to two-state memristor conductances and design layers including convolution, division, and signal conversion to achieve selective restoration, with an enhanced MSCE variant offering a 57.6% power reduction over MSC. Experimental results on grayscale images from BSD68 show MSC/MSCE achieving comparable PSNR/SSIM to a ternary baseline (TSC) up to 50% SAP density and surpassing it at 60% and higher, with MSCE providing the best performance at high density. The differential MSCE architecture and central-signal processing contribute to robustness against high-density SAP noise, while co-simulation with MATLAB and PSpice validates the circuit-level feasibility. The work demonstrates a promising hardware-efficient approach for high-density SAP noise removal with potential for low-power neuromorphic denoising in edge devices.

Abstract

In this article, we propose a memristor-based selective convolutional (MSC) circuit for salt-and-pepper (SAP) noise removal. We implement its algorithm using memristors in analog circuits. In experiments, we build the MSC model and benchmark it against a ternary selective convolutional (TSC) model. Results show that the MSC model effectively restores images corrupted by SAP noise, achieving similar performance to the TSC model in both quantitative measures and visual quality at noise densities of up to 50%. Note that at high noise densities, the performance of the MSC model even surpasses the theoretical benchmark of its corresponding TSC model. In addition, we propose an enhanced MSC (MSCE) model based on MSC, which reduces power consumption by 57.6% compared with the MSC model while improving performance.

Figures (9)

• Figure 1: The weight distributions of traditional convolutional layers (blue scatter points) and SeConv layers (red scatter points) in SeConvNet are shown in (a), while the distribution of weight counts and values is displayed through kernel density estimation plots located above and to the right, respectively. Normal distribution properties are observed in both traditional convolutional layers and SeConv layers. The proportion of weight ranges for SeConv layers and traditional convolutional layers is depicted in (b) and (c) respectively.
• Figure 2: Schematic of the SeConv model with a $3\times 3$ convolution kernel and a $5\times 5$ input tensor. (a) illustrates the denoising principle and process, where the preprocessed noisy image through the SeConv model yields the restored tensor. (b) shows the circuitry of the SeConv model, the input is kept consistent with (a), and the output results of both are compared, noting that the errors remain within an acceptable range. In (b), errors are denoted with blue markings.
• Figure 3: Simulation results of circuit in Fig. \ref{['fig_2']}(b).
• Figure 4: The circuit architecture of the MSCE model. The red and blue lines highlight the portions of the circuit that only select the central voltage of $\widetilde{M}_A$ and $\widetilde{A}$.
• Figure 5: Restoration results of different models. A1-A8 depict noisy images, while B1-B8, C1-C8, D1-D8, and E1-E8 represent restoration images by the FPSC, TSC, MSC, and MSCE models respectively.