应用空洞卷积的神经网络轻量化方法#br#

doi:10.3778/j.issn.1002-8331.2101-0020

摘要/Abstract

摘要： 为了深度卷积神经网络能够更好地应用于边缘设备，减少深度神经网络的模型参数量，降低网络复杂度，对于轻量化卷积神经网络的研究日益增多。将常用于多尺度融合的卷积——空洞卷积首次应用于神经网络轻量化领域研究。利用空洞卷积采样与扩大感受野的结构特性，实现了模型参数量的减少与计算复杂度的降低。将空洞卷积单纯作用于残差网络结构，达到轻量化目的，并进一步与逐点卷积结合，提高轻量化效果，形成改进型空洞卷积轻量化方法。为减少准确率衰减，将改进型空洞卷积与普通卷积相融合，提出一种融合型空洞卷积轻量化方法。实验结果表明，改进型空洞卷积轻量化方法具有最显著的轻量化效果，融合型空洞卷积轻量化方法使模型参数量减少同时具有最佳速度与精度的权衡。

关键词: 卷积神经网络, 轻量化, 空洞卷积, 融合, 精度与速度权衡

Abstract: In order to apply deep convolutional neural networks better to edge devices, reduce the amount of deep neural networks’ model parameters, and reduce network complexity, research on lightweight convolutional neural networks is increasing. In this article, some researches are first done in the field of lightweight neural networks about dilated convolution which is commonly used in multi-scale fusion. Using the structural characteristics of dilated convolution sampling and expanding the receptive field, the reduction of model parameters and the reduction of computational complexity are realized. Firstly, the dilated convolution is purely applied to the residual network structure to light convolutional neural networks, and further combined dilated convolution with the pointwise convolution to improve the lightweight effect, it proposes improved dilated convolution lightweight method. In order to reduce the attenuation of accuracy, it combines improved dilated convolution and ordinary convolution, and proposes a lightweight method of fusing dilated convolution. Experimental results show that the improved dilated convolution lightweight method has the most significant lightweight effect, the fusing dilated convolution lightweight method not only reduces the amount of model parameters, but also has the best speed and accuracy trade-off.

Key words: convolutional neural networks, lightweight method, dilated convolution, fusion, trade-off between accuracy and speed

马利, 刘新宇, 李皓宇, 段苛苛, 牛斌. 应用空洞卷积的神经网络轻量化方法#br#[J]. 计算机工程与应用, 2022, 58(5): 85-93.

MA Li, LIU Xinyu, LI Haoyu, DUAN Keke, NIU Bin. Neural Network Lightweight Method with Dilated Convolution[J]. Computer Engineering and Applications, 2022, 58(5): 85-93.

参考文献

[1] HOWARD A G，ZHU M，CHEN B，et al.MobileNets：efficient convolutional neural networks for mobile vision applications[J].arXiv：1704.04861，2017.
[2] SANDLER M，HOWARD A G，ZHU M，et al.MobileNetV2：inverted residuals and linear bottlenecks[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition，Salt Lake City，UT，USA，June 18-22，2018：4510-4520.
[3] IANDOLA F N，MOSKEWICZ M W，ASHRAF K，et al.SqueezeNet：AlexNet-level accuracy with 50x fewer parameters and <1MB model size[J].arXiv：1602.07360，2016.
[4] ZHANG X，ZHOU X，LIN M，et al.ShuffleNet：an extremely efficient convolutional neural network for mobile devices[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition，Salt Lake City，UT，USA，June 18-22，2018：6848-6856.
[5] MA N，ZHANG X，ZHENG H T，et al.ShuffleNet V2：practical guidelines for efficient CNN architecture design[C]//15th European Conference on Computer Vision，Munich，Germany，September 8-14，2018：122-138.
[6] HE Y，ZHANG X，SUN J.Channel pruning for accelerating very deep neural networks[C]//IEEE International Conference on Computer Vision，Venice，Italy，October 22-29，2017：1398-1406.
[7] HAN S，MAO H，DALLY W J.Deep compression：compressing deep neural network with pruning，trained quantization and huffman coding[C]//4th International Conference on Learning Representations，San Juan，Puerto Rico，May 2-4，2016.
[8] YIM J，JOO D，BAE J H，et al.A gift from knowledge distillation：fast optimization，network minimization and transfer learning[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition，Honolulu，HI，USA，July 21-26，2017：7130-7138.
[9] YU F，KOLTUN V.Multi-scale context aggregation by dilated convolutions[C]//4th International Conference on Learning Representations，San Juan，Puerto Rico，May 2-4，2016.
[10] CHEN L C，PAPANDREOU G，KOKKINOS I，et al.DeepLab：semantic image segmentation with deep convolutional nets，atrous convolution，and fully connected CRFs[J].IEEE Trans Pattern Anal Mach Intell，2018，40（4）：834-848.
[11] MEHTA S，RASTEGARI M，CASPI A，et al.ESPNet：efficient spatial pyramid of dilated convolutions for semantic segmentation[C]//15th European Conference on Computer Vision，Munich，Germany，September 8-14，2018：561-580.
[12] 毕鹏程，罗健欣，陈卫卫.轻量化卷积神经网络技术研究[J].计算机工程与应用，2019，55（16）：25-35.
BI Pengcheng，LUO Jianxin，CHEN Weiwei.Research on lightweight convolutional neural network technology[J].Computer Engineering and Applications，2019，55（16）：25-35.
[13] 葛道辉，李洪升，张亮，等.轻量级神经网络架构综述[J].软件学报，2020，31：2627-2653.
GE Daohui，LI Hongsheng，ZHANG Liang.et al.Survey of lightweight neural network[J].Journal of Software，2020，31：2627-2653.
[14] SZEGEDY C，VANHOUCKE V，IOFFE S，et al.Rethinking the inception architecture for computer vision[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition，Las Vegas，NV，USA，June 27-30，2016：2818-2826.
[15] LUO W，LI Y，URTASUN R，et al.Understanding the effective receptive field in deep convolutional neural networks[J].arXiv：1701.04128，2017.
[16] SIMONYAN K，ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[C]//3rd International Conference on Learning Representations，San Diego，CA，USA，May 7-9，2015.
[17] HE K，ZHANG X，REN S，et al.Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition，Las Vegas，NV，USA，June 27-30，2016：770-778.
[18] YU F，WANG D，DARRELL T.Deep layer aggregation[J].arXiv：1707.06484，2017.
[19] ZOU Z，SHI Z，GUO Y，et al.Object detection in 20 years：a survey[J].arXiv：1905.05055，2019.
[20] ZHOU X，WANG D，KR?HENBüHL P.Objects as points[J].arXiv：1904.07850，2019.
[21] LAW H，DENG J.CornerNet：detecting objects as paired keypoints[C]//15th European Conference on Computer Vision，Munich，Germany，September 8-14，2018：765-781.