Spiking Neural Network with Mixture of Heterogeneous Enhancement Experts for Robust Underwater Object Detection

Long Chen, Yaxin Li, Jiangrong Shen, Ling Huang, Yunzhi Zhuge, Qi Xu, Gang Pan, Qiang Zhang

Keywords: Spiking Neural Network, Mixture of Experts, Heterogeneous Enhancement Experts

Abstract: Underwater object detection faces unique challenges from haze, color distortion, and low contrast caused by light absorption and scattering, which significantly degrade image quality and detection performance. We propose HE-MoESNN, a spiking neural network that integrates a Mixture of Heterogeneous Enhancement Experts (HE-MoE) with a lightweight Forward Spiking Neural Network (FSNN) backbone. Unlike conventional MoE frameworks that feed identical inputs to all experts, HE-MoE assigns modality-specific inputs consisting of dehazing, color correction, and contrast enhancement to three parallel experts and fuses their outputs through a shared gating router. This design promotes expert diversity and enables the network to exploit complementary enhancement cues. FSNN improves efficiency by replacing costly ANN activations and conventional convolutions with signed spiking neurons and ternary convolutions, reducing computation while maintaining competitive accuracy. Extensive experiments on the RUOD and DUO benchmarks demonstrate that HE-MoESNN achieves state-of-the-art performance while maintaining high computational efficiency.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 9901