Neuromorphic adaptive edge-preserving denoising filter

TL;DR

This paper introduces a neuromorphic hardware implementation of an adaptive edge-preserving denoising filter that outperforms traditional filters in noise reduction while maintaining edge details.

Contribution

The paper proposes a novel neuromorphic spatial filter that adaptively preserves edges during denoising, using neuron-based similarity detection and hardware implementation.

Findings

Achieves higher PSNR and lower MSE than conventional filters.

Demonstrates robustness on noisy images from Caltech database.

Successfully preserves edges while reducing noise.

Abstract

In this paper, we present on-sensor neuromorphic vision hardware implementation of denoising spatial filter. The mean or median spatial filters with fixed window shape are known for its denoising ability, however, have the drawback of blurring the object edges. The effect of blurring increases with an increase in window size. To preserve the edge information, we propose an adaptive spatial filter that uses neuron's ability to detect similar pixels and calculates the mean. The analog input differences of neighborhood pixels are converted to the chain of pulses with voltage controlled oscillator and applied as neuron input. When the input pulses charge the neuron to equal or greater level than its threshold, the neuron will fire, and pixels are identified as similar. The sequence of the neuron's responses for pixels is stored in the serial-in-parallel-out shift register. The outputs of shift registers are used as input to the selector switches of an averaging circuit making this an adaptive mean operation resulting in an edge preserving mean filter. System level simulation of the hardware is conducted using 150 images from Caltech database with added Gaussian noise to test the robustness of edge-preserving and denoising ability of the proposed filter. Threshold values of the hardware neuron were adjusted so that the proposed edge-preserving spatial filter achieves optimal performance in terms of PSNR and MSE, and these results outperforms that of the conventional mean and median filters.