Layer-Wise Spike Modulation for Structure-Aware Image Denoising with Spiking Neural Networks

Noise arising from physical and environmental factors can degrade the accuracy of image analysis. This study proposes a spiking neural network (SNN)-based image denoising model that combines ResNet18-based feature extraction with LIF-based spike processing. The model was evaluated for Gaussian, shot, and horizontal and vertical stripe noise using standard image datasets and medical images. Experiments were also conducted under combined corruption conditions to assess restoration performance. The results showed that the configuration with LIF neurons after each ResNet18 block and a spike attenuation mechanism at Block 3 achieved the best restoration performance among the evaluated configurations. In particular, the proposed model achieved a PSNR improvement of up to 2.35 dB on the ISIC 2018 dataset for stripe noise removal. These findings suggest that layer-wise spike modulation can influence image denoising performance. Although the proposed model exhibited lower spike activity than the compared SNN model, its computational and energy-efficiency advantages require further evaluation on suitable neuromorphic hardware.