Deep Neural Network Model Compression And Its Application In Face Recognition

Deep learning has achieved great success in applications such as image,speech,and natural language processing,but deep neural networks models require a lot of computational and storage space overhead due to their complex network structure and a large number of network parameters.It is hard to deeploy on hardware devices with limited storage and computing resources,especially on mobile devices.Therefore,on the premise of ensuring that the accuracy of the deep neural networks model does not decrease,the network model can be streamlined to reduce the requirements on the computing capacity of hardware devices and the overhead of storage resources,which is an important research issue.This paper focuses on the deep neural networks model compression research,and applies it to the deep face network model,the specific work is as follows:1.Network pruning is a technology used to remove redundant connections of the network to reduce the amount of network redundancy calculation.The key of this technology is to define the network redundant connections,but at present,there is no accurate definition of network redundant connections.Although the weight of the network connection is related to the important connection of the network,it is necessary to determine the threshold of the important connection of the network.In order to solve the shortcomings of existing network pruning methods in determining network pruning thresholds and selecting important connection weights,this paper effectively proposes a new method to determine the network pruning threshold and select the important connection weight of the network based on the difference between the initial state and the final state of the network.2.Compared with the current face network models,which rely on the construction of complex network structures or non-public face data sets to achieve superior performance,this paper uses the residual network structure and MFM activation function to design a relatively small face network model in commonly used public face dataset and tested on the LFW face dataset to achieve higher accuracy.This paper constructs a new ResNet+MFM network module to improve the feature extraction capability of the network.In order to effectively improve the network accuracy,this paper uses the A-Softmax loss function to train the network model.3.In order to reduce the overhead of network model parameters on storage space,this paper combines the redundancy weight removal method with weight re-quantization and Huffman coding compression to compress the face network model of this paper to remove the redundant weight of the network model and reduce the storage space.