Attention-based Efficient Lightweight Model for Accurate Real-Time Face Verification on Embedded Device

With the rapid development of face verification technology, the current face verification has reached a high accuracy rate. However, face verification requires large computing resources and complex model parameters, so it is difficult to be applied in real scenes. The face verification system based on embedded applications can well solve the problems of huge model storage space occupation and high computing resource consumption. The existing mainstream face verification models have reached high accuracy, such as VGG16Net. Although there have been studies trying to design lightweight neural networks, such as MobileFaceNet, etc., there are still shortcomings such as complex structure and high resource consumption, so it is only suitable for mobile devices and other embedded devices, and it is difficult to apply to low power consumption and low performance such as embedded systems. In the device. For embedded systems with limited memory and computing power, a more lightweight neural network model is needed. Therefore, this paper conducts two researches on the problems of model storage requiring a large amount of space and high computational resource consumption. First of all, a lightweight face verification network model MobileFaceNet-v3m based on the attention mechanism is designed to solve the problem of model storage space occupation. Our model has a 15.26% smaller space occupation than MobileFaceNet, which successfully reduces the resource consumption of the model. The accuracy rate can still be maintained at a high level-95.47% of face verification accuracy is achieved on LFW; secondly, We use the learning rate optimization method based on efficient warm restart to train the model, the accuracy of the model is improved faster, and the model is successfully transplanted to the embedded platform, which is another step forward to the real scene application in the future.