Pruning rPPG Networks: Toward Small Dense Network with Limited Number of Training Samples

Neural network pruning reduces network complexity and storage by removing unimportant connections in the network, enabling network miniaturization, fast training and inference, easy deployment to portable devices, etc. The emerging lottery ticket hypotheses and sparse initialization technique have shed new lights on the pruning research. However, few research focuses on the pruning of the networks for remote photoplethysmography (rPPG) pulse signal extraction. Opposite to the existing pruning researches that prune large network, rPPG networks are relatively small. It is interesting to see how it behaves when the pruning is applied. In this paper, we investigate the behavior of common pruning techniques when applied to an existing rPPG network. Experiments on PURE dataset show that the pruning rate decay is beneficial to the performance improvement, whereas the connection regeneration has a detrimental effect. Given the same final sparsity, dense initialization generally performs better than sparse initialization. The network seems insensitive to initial sparsity. The combination s
i
=1.0, s
f
=0.1, with decay, and without regeneration is the best trade-off between SNR and FLOPs, achieving average SNR 9.78 dB, increased by 0.48 dB in comparison with the original PhysNet.