Energy-efficient face detection and recognition based on ANN–SNN conversion

Abstract Spiking neural networks (SNNs) exhibit lower power consumption and faster inference speeds compared with artificial neural networks (ANNs), garnering significant research interest. ANN–SNN conversion is a popular method for object detection, image recognition, etc. Designing an architecture for image detection and recognition that balances high accuracy and low power consumption via ANN–SNN conversion remains a critical research challenge. In this paper, we propose a module called downward feature pyramid network to enhance the correlation between feature maps of different scales among networks. In response to the incompatibility between the efficient modules of ANN and the sparsity of SNN, we construct an SNN face detection and recognition framework called SNN-FDR based on RetinaFace and AdaFace. We conduct extensive experiments on the datasets WIDER FACE, LFW, AgeDB, CALFW, etc. Experimental results demonstrate a 95% reduction in energy consumption during the detection phase compared with the baseline, with only a 4.7% accuracy degradation. And in the recognition stage, we reduced energy consumption by 44.8% on high-quality datasets, while the average accuracy decreased only by 1.5%.