SNNF: An SNN-based Near-Sensor Noise Filter for Dynamic Vision Sensors
Spiking neural network filter achieves 44.4 Meps with <5% power of ANN designs
Dynamic Vision Sensors (DVS) are ideal for edge IoVT applications due to low power and high dynamic range, but their output suffers from spurious Background Activity (BA) noise that wastes compute. A team of researchers has now published SNNF (SNN-based Near-Sensor Noise Filter), a hardware-efficient solution that filters BA noise using a single-layer Spiking Neural Network (SNN) classifier. The design introduces a compact Event-Based Binary Image (EBBI) representation that removes timestamp dependency, drastically cutting memory footprint. The SNN classifier achieves an AUC of 0.89 on standard datasets, distinguishing signal events from noise. By replacing power-hungry multipliers with simple accumulation logic and minimizing inter-neuron data width, the filter is extremely hardware-efficient.
FPGA implementation results show SNNF uses only about 11% of the memory and 40% of the logic resources compared to state-of-the-art filters, while delivering a throughput of 29 Mega events per second (Meps). In a 65 nm CMOS ASIC implementation, performance scales to 44.4 Meps with area and power consumption at just ~13% and <5% of equivalent ANN-based designs. These results demonstrate that SNNF achieves an excellent balance between filtering accuracy and hardware efficiency, making it highly suitable for resource-constrained, near-sensor deployment in real-time vision systems.
- SNNF uses a single-layer Spiking Neural Network classifier with 0.89 AUC to filter Background Activity noise in DVS sensors.
- FPGA implementation achieves 29 Meps throughput using only 11% memory and 40% logic of prior state-of-the-art filters.
- In 65nm CMOS ASIC, SNNF delivers 44.4 Meps with 13% area and under 5% power consumption vs ANN-based designs.
Why It Matters
Enables efficient near-sensor vision processing for IoVT edge devices with minimal compute and battery drain.