解耦同类自知识蒸馏的轻量化唇语识别方法

马金林; 刘宇灏; 马自萍; 郭兆伟; 吕鑫

doi:10.13700/j.bh.1001-5965.2022.0931

解耦同类自知识蒸馏的轻量化唇语识别方法

doi: 10.13700/j.bh.1001-5965.2022.0931

马金林^{1, 2, ,},
刘宇灏¹,
马自萍³,
郭兆伟¹,
吕鑫¹

1.
北方民族大学计算机科学与工程学院，银川 750021
2.
北方民族大学图像图形智能处理国家民委重点实验室，银川 750021
3.
北方民族大学数学与信息科学学院，银川 750021

基金项目: 宁夏自然科学基金(2022AAC03268)；北方民族大学中央高校基本科研业务费专项资金(2021KJCX09,FWNX21)；北方民族大学“计算机视觉与虚拟现实”创新团队项目

详细信息

通讯作者:
E-mail：majinlin@nmu.edu.cn

中图分类号: TP391
计量
- 文章访问数: 299
- HTML全文浏览量: 101
- PDF下载量: 7
- 被引次数: 0
出版历程
- 收稿日期: 2022-11-17
- 录用日期: 2023-02-17
- 网络出版日期: 2023-03-10
- 整期出版日期: 2024-12-31

Lightweight lip reading method based on decoupling homogeneous self-knowledge distillation

MA Jinlin^{1, 2
, ,},
LIU Yuhao¹,
MA Ziping³,
GUO Zhaowei¹,
LYU Xin¹

1.
School of Computer Science and Engineering，North Minzu University，Yinchuan 750021，China
2.
Key Laboratory of Image and Graphics Intelligent Processing of State Ethnic Affairs Commission，North Minzu University，Yinchuan 750021，China
3.
School of Mathematics and Information Science，North Minzu University，Yinchuan 750021，China

Funds: Natural Science Foundation of Ningxia Autonomous Region (2022AAC03268); Basic Research Funds for Central Universities of North Minzu University (2021KJCX09,FWNX21); The Innovation Team of Computer Vision and Virtual Reality of North Minzu University

More Information

Corresponding author: E-mail：majinlin@nmu.edu.cn

摘要

摘要:
针对唇语识别模型因参数量和计算量较大而无法在移动终端和边缘设备上应用的问题，提出基于解耦同类自知识蒸馏和GhostNet-TSM的唇语识别方法。提出具有时序特征提取能力的GhostNet-TSM网络；将同类自知识蒸馏的特征解耦为目标类特征和非目标类特征，分别设置损失函数，以提高模型的识别精度；使用解耦同类自知识蒸馏方法在LRW和LIP350数据集上进行模型训练，并在OuluVS数据集上进行验证。实验结果表明：GhostNet-TSM网络在LRW数据集上达到了85.2％的识别准确率，超过了多数非轻量化模型，浮点数计算量和模型参数量降低至0.988 GFLOPs和20.310×10⁶。
- 唇语识别 /
- 知识蒸馏 /
- 轻量化 /
- GhostNet /
- TSM模块
Abstract:
Due to its huge parameters and computational requirements, the lip reading model is not suitable for use with edge devices or mobile terminals. In order to solve this problem, we propose a lip reading method based on decoupled homogeneous self-knowledge distillation and GhostNet-TSM. First, the TSM module is installed into the GhostNet to extract temporal features. Second, the homogenous self-knowledge distillation is decoupled to increase the recognition accuracy of the model. Finally, the validation experiment on the LRW and LIP350 datasets has been completed and verified on the OuluVS dataset.The experimental results show that the recognition accuracy of GhostNet-TSM on the LRW dataset reaches 85.2%, which exceeds that of most non-lightweight models. The number of floating point operations per second and model parameters is reduced to 0.988 GFLOPs and 20.310×10⁶.
- lip reading /
- knowledge distillation /
- lightweight /
- GhostNet /
- TSM module

HTML全文

图 1 普通卷积与Ghost模块

Figure 1. Convolution and Ghost module

下载: 全尺寸图片幻灯片

图 2 Ghost瓶颈模块结构

Figure 2. Structure of Ghost bottleneck module

下载: 全尺寸图片幻灯片

图 3 普通知识蒸馏原理

Figure 3. Principles of general knowledge distillation

下载: 全尺寸图片幻灯片

图 4 同类自知识蒸馏原理

Figure 4. Principles of homogeneous self-knowledge distillation

下载: 全尺寸图片幻灯片

图 5 TSM模块原理

Figure 5. Principles of TSM

下载: 全尺寸图片幻灯片

图 6 GhostNet-TSM结构

Figure 6. Structure of GhostNet-TSM

下载: 全尺寸图片幻灯片

图 7 TSM模块的2种插入方式

Figure 7. Two kinds of inserting TSM module

下载: 全尺寸图片幻灯片

图 8 同类自知识蒸馏与解耦后的流程

Figure 8. Decoupling homogeneous self-knowledge distillation and decoupled process

下载: 全尺寸图片幻灯片

图 9 LIP350数据采集示例

Figure 9. Samples of LIP350 data acquisition

下载: 全尺寸图片幻灯片

表 1 T和N对识别性能的影响

Table 1. Influence of T and N on recognition performance

网络	蒸馏或正则化方法	T	N	准确率/%	变化量/%
GhostNet-TSM	无蒸馏			79.3
	同类自知识蒸馏	√	√	80.2	+0.9
	同类自知识蒸馏（仅T）	√		77.6	−1.7
	同类自知识蒸馏（仅N）		√	80.7	+1.4
MobileNet-TSM	无蒸馏			75.3
	同类自知识蒸馏	√	√	76.1	+0.8
	同类自知识蒸馏（仅T）	√		74.4	−0.9
	同类自知识蒸馏（仅N）		√	76.3	+1.0
ResNet101-TSM	无蒸馏			83.6
	同类自知识蒸馏	√	√	84.5	+0.9
	同类自知识蒸馏（仅T）	√		83.0	−0.6
	同类自知识蒸馏（仅N）		√	84.4	+0.8

下载: 导出CSV

表 2 不同网络结合TSM模块的对比实验结果

Table 2. Comparative experiment of different networks combining TSM modules

网络组成	基础网络	网络类型	准确率/%	参数量	浮点数运算量/GFLOPS
ResNet18-TSM	ResNet18	非轻量级	74.9	11.356×10⁶	7.320
ResNet34-TSM	ResNet34	非轻量级	79.8	21.464×10⁶	14.995
ResNet101-TSM	ResNet101	非轻量级	83.6	43.217×10⁶	33.180
DenseNet121-TSM	DenseNet121	非轻量级	77.6	51.182×10⁶	37.976
MobileNet-TSM	MobileNetV2	轻量级	75.3	3.672×10⁶	1.300
ShuffleNet v2-TSM	ShuffleNet V2	轻量级	71.1	4.823×10⁶	1.492
GhostNet-TSM (本文)	GhostNet	轻量级	79.3	4.350×10⁶	0.678266

下载: 导出CSV

表 3 与其他蒸馏方法的性能对比

Table 3. Performance comparison with other distillation methods

蒸馏或正则化方法	准确率/%
无蒸馏	79.3
标签平滑	79.5
普通知识蒸馏	77.6
自知识蒸馏	79.7
同类自知识蒸馏	80.2
解耦同类自知识蒸馏（仅使用N）	80.7

下载: 导出CSV

表 4 解耦参数对比

Table 4. Comparison of parameters of decoupling

T与N的系数	准确率/%
T与N的系数	LIP350	LRW
$ \alpha $= 0, $ \beta $= 0.8	80.4	82.0
$ \alpha $= 0, $ \beta $= 1	80.7	82.5
$ \alpha $= 0, $ \beta $= 2	80.9	82.1
$ \alpha $= 0, $ \beta $= 4	81.5	83.6
$ \alpha $= 0, $ \beta $= 5	81.3	82.6
$ \alpha $= 0, $ \beta $= 6	81.1	83.2
$ \alpha $= 0, $ \beta $= 7	79.5	81.4
$ \alpha $= 0, $ \beta $= 8	77.1	80.8
$ \alpha $= 0, $ \beta $= 4	81.5	83.6
$ \alpha $= 0.2, $ \beta $= 4	82.1	83.9
$ \alpha $= 0.3, $ \beta $= 4	82.5	83.3
$ \alpha $= 0.4, $ \beta $= 4	83.3	84.5
$ \alpha $= 0.5, $ \beta $= 4	82.9	84.6
$ \alpha $= 0.6, $ \beta $= 4	82.8	85.2
$ \alpha $= 0.8, $ \beta $= 4	81.3	84.8
$ \alpha $= 1, $ \beta $= 4	80.8	83.5

下载: 导出CSV

表 5 OuluVS数据集上验证实验结果

Table 5. Validation experiment on OuluVS dataset

说话者	蒸馏方法	准确率/%
7号	无蒸馏	91.2
	同类自知识蒸馏	91.8
	解耦同类自知识蒸馏	94.6
9号	无蒸馏	81.0
	同类自知识蒸馏	82.3
	解耦同类自知识蒸馏	86.4
14号	无蒸馏	87.5
	同类自知识蒸馏	88.3
	解耦同类自知识蒸馏	89.8

下载: 导出CSV

表 6 与其他唇语识别方法的性能对比

Table 6. Performance comparison with other lip recognition methods

唇语识别方法	年份	唇语识别模型的网络组成		参数量	浮点数运算量/GFLOPs	准确率/%
唇语识别方法	年份	前端	后端	参数量	浮点数运算量/GFLOPs	准确率/%
文献[30]	2017	VGG-M				61.1
文献[31]	2017	VGG-M	LSTM	419.256×10⁶	32.744	76.2
文献[32]	2019	ResNet34 + DenseNet3D	Conv-BLSTM	331.603×10⁶	29.386	83.3
文献[10]	2017	3D-Conv + ResNet34	BLSTM	93.168×10⁶	11.539	83.0
文献[27]	2018	3D-Conv + ResNet34	BLSTM	132.581×10⁶	13.681	83.4
文献[33]	2019	(3D-Conv)×2	BLSTM	976.867×10⁶	47.806	84.1
文献[26]	2018	3D-Conv + ResNet18	BLSTM	55.378×10⁶	9.332	84.3
文献[29]	2020	3D-Conv + ResNet18	BGRU	57.644×10⁶	10.055	84.4
文献[28]	2020	3D-Conv + ResNet18	BGRU	59.520×10⁶	10.598	85.0
文献[34]	2020	3D-Conv+P3D-ResNet	TCN			84.8
文献[35]	2020	SpotFast	TransformerProduct-Key memory			84.4
文献[13]	2020	3D-Conv + ResNet18	MS-TCN	52.872×10⁶	9.120	85.3
文献[36]	2021	3D-Conv + ResNet18	BGRU+ Visual-Audio Memory			85.4
文献[37]	2022	MoCo + Wav2Vec by SJTU LUMIA	MoCo + Wav2Vec by SJTU LUMIA			85.0
本文方法		GhostNet+TSM+解耦同类自知识蒸馏	GhostNet+TSM+解耦同类自知识蒸馏	20.310×10⁶	0.988	85.2

下载: 导出CSV

参考文献(37)

[1]	LESANI F S, GHAZVINI F F, DIANAT R. Mobile phone security using automatic lip reading[C]//Proceedings of the 9th International Conference on e-Commerce in Developing Countries: With Focus on e-Business. Piscataway: IEEE Press, 2015.
[2]	MATHULAPRANGSAN S, WANG C Y, KUSUM A Z, et al. A survey of visual lip reading and lip-password verification[C]//Proceedings of the International Conference on Orange Technologies. Piscataway: IEEE Press, 2015: 22-25.
[3]	BAHDANAU D, CHOROWSKI J, SERDYUK D, et al. End-to-end attention-based large vocabulary speech recognition[C]//Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE Press, 2016: 4945-4949.
[4]	HUANG J T, LI J Y, GONG Y F. An analysis of convolutional neural networks for speech recognition[C]//Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE Press, 2015: 4989-4993.
[5]	CHAE H, KANG C M, KIM B, et al. Autonomous braking system via deep reinforcement learning[C]//Proceedings of the IEEE 20th International Conference on Intelligent Transportation Systems. Piscataway: IEEE Press, 2017: 1-6.
[6]	HONG X P, YAO H X, WAN Y Q, et al. A PCA based visual DCT feature extraction method for lip-reading[C]//Proceedings of the International Conference on Intelligent Information Hiding and Multimedia. Piscataway: IEEEPress, 2006: 321-326.
[7]	PUVIARASAN N, PALANIVEL S. Lip reading of hearing impaired persons using HMM[J]. Expert Systems with Applications, 2021, 57(24): 61-73.
[8]	马金林, 朱艳彬, 马自萍, 等.唇语识别的深度学习方法综述[J]. 计算机工程与应用, 2021, 57(24): 61-73. MA J L,ZHU Y B,MA Z P,et al. Survey of deep learning methods for lip recognition[J]. Computer Engineering and Applications, 2021, 57(24): 61-73(in Chinese).
[9]	马金林, 陈德光, 郭贝贝, 等. 唇语语料库综述[J]. 计算机工程与应用, 2019, 55(22): 1-13. MA J L, CHEN D G, GUO B B,et al. Lip corpus review[J]. Computer Engineering and Applications, 2019, 55(22): 1-13(in Chinese).
[10]	STAFYLAKIS T, TZIMIROPOULOS G. Combining residual networks with LSTMs for lipreading[C]//Proceedings of the 18th Annual Conference of the International Speech Communication Association. Toulouse: ISCA-INT Speech Communication Association, 2017: 3652-3656.
[11]	ASSAELY M, SHILLINGFORD B, WHITESON S, et al. LipNet: End-to-end sentence-level lipreading[J]. (2016-12-16)[2022-11-01].
[12]	YANG H, YUAN C F, ZHANG L, et al. STA-CNN: Convolutional spatial-temporal attention learning for action recognition[J]. IEEE Transactions on Image Processing, 2020, 29: 5783-5793. doi: 10.1109/TIP.2020.2984904
[13]	MARTINEZ B, MA P C, PETRIDIS S, et al. Lipreading using temporal convolutional networks[C]//Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE Press, 2020: 6319-6323.
[14]	HAN K, WANG Y H, TIAN Q, et al. GhostNet: More features from cheap operations[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2020: 1577-1586.
[15]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2016: 770-778.
[16]	HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[EB/OL]. (2015-05-09)[2022-11-01].
[17]	LIU Y F, SHU C Y, WANG J D, et al. Structured knowledge distillation for dense prediction[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(6): 7035-7049. doi: 10.1109/TPAMI.2020.3001940
[18]	YANG Z D, LI Z, JIANG X H, et al. Focal and global knowledge distillation for detectors[EB/OL]. (2022-05-09)[2022-11-01].
[19]	MOBAHI H, FARAJTABAR M, BARTLETT P L. Self-distillation amplifies regularization in Hilbert space[EB/OL]. (2020-10-26) [2022-11-01].
[20]	ZHANG Z L, SABUNCU M R. Self-distillation as instance-specific label smoothing[EB/OL]. (2020-10-22)[2022-11-01].
[21]	YUN S, PARK J, LEE K, et al. Regularizing class-wise predictions via self-knowledge distillation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2020: 13873-13882.
[22]	LIN J, GAN C, HAN S. TSM: Temporal shift module for efficient video understanding[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE Press, 2019: 7082-7092.
[23]	KING D E. Dlib-ml: A machine learning toolkit[J]. Journal of Machine Learning Research, 2009, 10: 1755-1758.
[24]	SANDLER M, HOWARD A, ZHU M L, et al. MobileNetV2: Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2018: 4510-4520.
[25]	MA N N, ZHANG X Y, ZHENG H T, et al. ShuffleNet V2: Practical guidelines for efficient CNN architecture design[C]//Proceedings of the European Conference on Computer Vision. Berlin: Springer, 2018: 122-138.
[26]	STAFYLAKIS T, KHAN M H, TZIMIROPOULOS G. Pushing the boundaries of audiovisual word recognition using Residual Networks and LSTMs[J]. Computer Vision and Image Understanding, 2018, 176-177: 22-32. doi: 10.1016/j.cviu.2018.10.003
[27]	PETRIDIS S, STAFYLAKIS T, MA P C, et al. Audio-visual speech recognition with a hybrid CTC/attention architecture[C]//Proceedings of the IEEE Spoken Language Technology Workshop. Piscataway: IEEE Press, 2018: 513-520.
[28]	ZHANG Y H, YANG S, XIAO J Y, et al. Can we read speech beyond the lips? Rethinking RoI selection for deep visual speech recognition[C]//Proceedings of the 15th IEEE International Conference on Automatic Face and Gesture Recognition. Piscataway: IEEE Press, 2020: 356-363.
[29]	ZHAO X, YANG S, SHAN S G, et al. Mutual information maximization for effective lip reading[C]//Proceedings of the 15th IEEE International Conference on Automatic Face and Gesture Recognition. Piscataway: IEEE Press, 2020: 420-427.
[30]	CHUNG J S, ZISSERMAN A. Lip reading in the wild[C]//Proceedings of the European Conference on Computer Vision. Berlin: Springer, 2017: 87-103.
[31]	CHUNG J S, SENIOR A, VINYALS O, et al. Lip reading sentences in the wild[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 3444-3453.
[32]	WANG C H. Multi-grained spatio-temporal modeling for lip-reading[EB/OL]. (2019-09-02)[2022-11-01].
[33]	WENG X S, KITANI K. Learning spatio-temporal features with two-stream deep 3D CNNs for lipreading[EB/OL]. (2019-07-19)[2022-11-01].
[34]	XU B, LU C, GUO Y D, et al. Discriminative multi-modality speech recognition[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2020: 14421-14430.
[35]	WIRIYATHAMMABHUM P. SpotFast networks with memory augmented lateral transformers for lipreading[C]//Proceedings of the International Conference on Neural Information Processing. Berlin: Springer, 2020: 554-561.
[36]	KIM M, HONG J, PARK S J, et al. Multi-modality associative bridging through memory: speech sound recollected from face video[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE Press, 2021: 296-306.
[37]	PAN X C, CHEN P Y, GONG Y C, et al. Leveraging unimodal self-supervised learning for multimodal audio-visual speech recognition[EB/OL]. (2022-05-26)[2022-11-01].