基于快速下采样的轻量化网络设计方法及人脸识别应用

High-precision convolutional neural networks often come with high inference costs, making it difficult to perform real-time inference on resource-constrained embedded devices. We analyze the factors that influence the speed of model inference by different types of convolutions, and for the first time point out that in addition to the computational complexity of the model, the feature map throughput of the model is also a key factor affecting the inference speed. However, the existing lightweight methods based on the depth-wise separation convolution only use computational complexity as the model lightweight metric, not considering the influence of the feature map throughput on the model inference speed. Based on this discovery, we propose a model lightweight acceleration design method combined with standard convolution based on fast down-sampling module, which could reduce the computational complexity and feature map throughput of the model at the same time by rapidly reducing the size of the feature map. The performance and the inference speed on different platforms of the models designed by proposed method are better than the existing lightweight methods based on depth-wise separation convolution. Further, we utilize this method to propose a fast face recognition model FDFaceNet(Fast Down-sampling FaceNet)for face recognition tasks. Compared with the existing lightweight face recognition models, FDFaceNet has higher accuracy and faster inference speed on various platforms.