Enhancing MobileNetV2 Performance with Layer Replication and Splitting for 3D Face Recognition Task Using Distributed Training

In this paper, we propose the operators, which are layer replication and splitting, to modify the CNN network. The algorithm for exploring suitable modifications to the prototype MobileNetV2 architecture for the 3D face recognition task is presented. The algorithm consists of two steps. The first step involves splitting the input and searching for the optimal position to expand the network through layer replication. The second part explores the architecture modification by altering the split input layer position within the model. Experiments demonstrate that the discovered model provides a more cost-effective performance, with only 0.015% increased size compared to a vanilla MobileNet V2, while delivering 6.99% higher accuracy than the previous 3D MobileNetV2 and 9.15% more accuracy than the vanilla MobileNetV2, with a total of 3,801,136 parameters. Using distributed data training on multi-GPUs, the total training time can be reduced by 75% while maintaining good accuracy compared to traditional single-GPU training.