SNN-Based Semantic Segmentation Method Using Adaptive Threshold and Multi-Feature Fusion

Spiking neural networks (SNN) are asynchronous and sparse, making them a low-power alternative to artificial neural networks (ANN) in specific scenarios. However, SNNs exhibit deficiencies such as limited feature expression and low result accuracy when utilized for complex tasks such as semantic segmentation. To address this issue, this paper proposes a SNN-based semantic segmentation methodology using adaptive thresholds for spike triggering and multi-feature fusion. First, input images are encoded utilizing adaptive threshold neurons for spiking, and the spiking visual geometry group (VGG) is utilized for spiking feature extraction from the encoded results. To further improve the network's ability of feature extraction for spike sequences, a spiking feature fusion (SFF) module designed at the encoder end subsequently functions to perform multi-scale feature extraction and fusion of the spike sequences. In the final decoding stage, the spike sequences are stacked in the time dimension, and spiking features at the encoding and decoding layers are fused to improve the network's capability in boundary feature extraction. Experimental results showed that the proposed approach achieved a mean intersection of union (mIoU) on the VOC2012 dataset that was 6.8% higher than that of existing SNN semantic segmentation algorithms. This solution also yielded good segmentation effect on the OCS dataset in high-speed railway scenarios. Additionally, the proposed method consumed less energy compared to ANN-based semantic segmentation algorithms.