SenSNN: Energy-Efficient Tiny-Object Pseudo-Mask Segmentation via a Dual-Stream Membrane Shortcut

Abstract

Real-time monitoring on nanosatellites is constrained by limited power and computation. While convolutional neural networks achieve strong performance in terrestrial vision tasks, their dense multiply-accumulate (MAC) operations make them costly for continuous onboard screening. To address this challenge, we propose SenSNN, a hardware-aware spiking neural network for remote-sensing tiny-object pseudo-mask segmentation. SenSNN introduces a dual-stream membrane shortcut to preserve subthreshold responses of microtargets and mitigate feature decay in deep spiking layers. It also employs a localized multiscale temporal consensus module to suppress redundant background spikes and improve the accuracy–energy tradeoff. We analyze SenSNN through a Pareto-oriented design space study, showing that localized spiking consensus provides a more efficient tradeoff than global attention for subpixel targets. Since SODA-A provides instance-level polygon annotations rather than native semantic segmentation masks, all SODA-A experiments in this work are conducted under a polygon-derived binary foreground pseudo-mask segmentation protocol. Under this protocol, SenSNN achieves 21.87% mIoU in its high-accuracy configuration. Its efficiency-oriented configuration achieves 20.87% mIoU and is estimated to require 4.18 mJ per inference under a 45-nm operation-level MAC/accumulate energy model, providing an approximately 4× estimated arithmetic-energy reduction compared with the resolution-matched MobileNetV2-UNet baseline (20.83% mIoU and 16.9 mJ). Additional supervised adaptation experiments on AI-TOD and DOTA-v1.0 further support the applicability of SenSNN on independent remote-sensing benchmarks. These results indicate that SenSNN provides a complementary low-energy architecture for dense tiny-object screening under constrained operation-level energy budgets.