Adaptive high-precision compression method and system based on convolutional neural network model

Patent metadata

Field	Value
Publication number	US-12380331-B2
Application number	US-202117448934-A
Country	US
Kind code	B2
Filing date	Sep 27, 2021
Priority date	Apr 30, 2021
Publication date	Aug 5, 2025
Grant date	Aug 5, 2025

Abstract

The present disclosure discloses an adaptive high-precision compression method and system based on a convolutional neural network model, and belongs to the fields of artificial intelligence, computer vision, and image processing. According to the method of the present disclosure, coarse-grained pruning is performed on a neural network model by using a differential evolution algorithm first, and the coarse-grained space is quickly searched through an entropy importance criterion and an objective function with good guidance to obtain a near-optimal neural network structure. Then fine-grained search space is built on the basis of an optimal individual obtained from the coarse-grained search, and fine-grained pruning is performed on the neural network model by a differential evolution algorithm to obtain a network model with an optimal structure. Finally, the performance of the optimal model is restored by using a multi-teacher multi-step knowledge distillation network to reach the precision of an original model.

First claim

What is claimed is:

1. A facial expression recognition method, the method comprising:
obtaining sample images, marking facial expressions to be recognized in the sample images, and dividing the sample images into a train set and a test set;
performing scaling, rotation, cropping, and normalization operations on the sample images, and then unify image sizes;
building a facial expression recognition neural network, wherein the neural network has more than 40 layers, and a neural network combining VGG and Resnet is used;
training the neural network built in the building step with the sample images obtained in the performing step to obtain an original neural network;
compressing the original neural network obtained in the training step; and
performing facial expression recognition with the neural network that has been compressed in the compressing step;
wherein compressing the original neural network obtained in the training step comprises:
training an original model on a train data set, and calculating importance of each channel of each layer of the original model on an importance evaluation data set;
performing structure search pruning in coarse-grained search space to obtain a near-optimal model structure, according to a designed differential evolution pruning algorithm;
performing fine-grained structure search on a basis of coarse-grained pruning to obtain an optimal model structure, according to the designed differential evolution pruning algorithm; and
performing knowledge distillation on the optimal model structure obtained by pruning to obtain a compression model with improved precision, according to a designed multi-step multi-teacher distillation method;

wherein performing fine-grained structure search on the basis of coarse-grained pruning to obtain the optimal model structure, according to the designed differential evolution pruning algorithm comprises:
reducing the range of the search space, and setting both a search range granularity and a change scale in a variation process to 1,

wherein the search space of fine-grained pruning is to select channel structures of optimum three individual in coarse-grained pruning to rebuild, and find the maximum value and minimum value of the number of channels in each layer of the three structures as a search range; and

optimum three models of coarse-grained pruning are selected as initial parents of fine-grained pruning, and then adaptive structure search is performed in the search space of fine-grained pruning to obtain an optimal model.

2. The facial expression recognition method according to claim 1 , wherein training the original model on the train data set, and calculating importance of each channel of each layer of the original model on the importance evaluation data set comprises:
selecting an entropy as an evaluation criterion for the importance of the channel; the entropy of each output channel of each layer of an original model is calculated through an inference process, and the definition of the channel entropy is as shown in Formula (1);
H=Σ i=0 255 p i log p i   (1)

wherein p i represents a ratio of the number of values within a range of [i,i+1] in the channel to the total number of values in the channel.

3. The facial expression recognition method according to claim 1 , wherein performing structure search pruning in coarse-grained search space to obtain the near-optimal model structure, according to the designed differential evolution pruning algorithm comprises:
S11: setting the maximum value and the minimum value of search space for each layer of neural network according to a data set and prior knowledge, and making 4 equal divisions, or 8 equal divisions, or 16 equal divisions within the set value range,
where the form of the search space is coarse_space=[c 1 , c 2 , . . . , c i , . . . , c n ], c i represents the search range of the ith layer, with a form of c i =[c i1 , c i2 , . . . , c ij , . . . , c im ], and c ij represents the number of the jth channels in the ith layer;
S12: setting an objective function as shown in Formula (2):

f
⁡
(
C
)

=

argmin
C

(

α
1

(

1
-

acc
⁡
(

N
⁡
(

C
,

D
train

)

,

D
test

)

+

α
2

⁢



P
cur





P
org



+

α
3

⁢

M
cur

M
org

)

)

(
2
)

where acc(N(C,D train ), D test ) represents the accuracy of a pruning network with a channel structure C tested on a test data set D test after being trained on a train data set D train ; M cur and P cur respectively represent the number of computational operations and the number of parameters of a model with a channel structure C, M org and P org

Assignees

• Univ Chongqing
• Univ Electronic Sci & Tech China
• Dibi Chongqing Intelligent Tech Research Institute Co Ltd
• Star Inst Of Intelligent Systems

Inventors

• Song Yongduan
• Yang Feng
• Li Rui
• Pan Shengtao
• Li Siyu
• Zhang Yiwen
• Zhang Jian
• Yu Zhengtao
• Wang Shichun

Classifications

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/082 Primary

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464 Primary

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/045 Primary

  Combinations of networks · CPC title