基于多分支瓶颈结构的轻量型图像分类算法研究

doi:10.12263/DZXB.20220920

摘要/Abstract

摘要：

传统卷积神经网络存在着参数量大、训练耗时长、轻量级模型的识别准确度不足的问题.本文提出了一种基于ResNet网络的多分支结构轻量化网络（Residual multi-branch structured Network，RemulbNet），通过在残差结构的主干中使用多分支结构增加特征多样性，利用变体的深度可分离卷积缩减模型参数量，采用Mish激活函数增加网络的非线性表达能力，在有效减少模型体积的情况下，提升网络的分类准确率.利用图像识别数据库，对网络性能进行测试.研究表明，对于5分类花卉识别问题，RemulbNet相比ResNet网络识别准确率提高3.9%，模型参数量减小71%，模型体积减小77%，缩短了约40%训练耗时；与轻量级网络（MobileNet v2和ShuffleNet v2）相比，RemulbNet在识别准确度、模型体积、训练时长和不同的图像分类数据集上都表现出优良的性能.

关键词: 轻量化网络, 多分支瓶颈结构, Mish激活函数, 深度可分离卷积, 图像分类, 卷积神经网络

Abstract:

The traditional convolutional neural networks have many problems, such as large number of parameters, long training time, and insufficient recognition accuracy of the lightweight models. Based on ResNet network, a lightweight network named RemulbNet(Residual multi-branch structured Network) with multi-branch structure, which increases feature diversity with multi-branch structure in the backbone of the residual structure, reduces the number of model parameters with the depth-separable convolution of variants, and also increases the nonlinear expression capability of the network with Mish activation function. These measures can effectively reduce the model volume and improve the classification accuracy of the network. Using the image recognition database, the network performance is tested. For 5 categories of flower identification, RemulbNet improves the recognition accuracy by 3.9%, reduces the number of model parameters by 71%, reduces the model volume by 77%, and shortens the training time by about 40% compared with the ResNet network. Facing different image classification datasets, RemulbNet also shows excellent performance in terms of recognition accuracy, model volume, training time compared with the lightweight networks(MobileNet v2 and ShuffleNet v2).

Key words: lightweight networks, multi-branch bottleneck structure, mish activation function, deep separable convolution, image classification, convolutional neural network

中图分类号:

TP391

苏田田, 王慧敏, 张小凤. 基于多分支瓶颈结构的轻量型图像分类算法研究[J]. 电子学报, DOI: 10.12263/DZXB.20220920.

SU Tian-tian, WANG Hui-min, ZHANG Xiao-feng. Research on Lightweight Image Classification Algorithm Based on Multi-branch Bottleneck Structure[J]. Acta Electronica Sinica, DOI: 10.12263/DZXB.20220920.

图/表 13

参考文献 22

1	FUKUSHIMA K. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position[J]. Biological Cybernetics, 1980, 36(4): 193-202.
2	LECUN Y, BOTTOU L. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
3	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[C]//Proceedings of the 25th International Conference on Neural Information Processing Systems-Volume 1. Lake Tahoe: MIT Press, 2012: 1106-1114.
4	TAJBAKHSH N, SHIN J Y, GURUDU S R, et al. Convolutional neural networks for medical image analysis: Full training or fine tuning?[J]. IEEE Transactions on Medical Imaging, 2016, 35(5): 1299-1312.
5	SZEGEDY C, LIU W, JIA Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston: IEEE, 2015: 1-9.
6	SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[EB/OL]. (2014-09-04)[2022-08-05]. .
7	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 770-778.
8	IOFFE S, SZEGEDY C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[C]//International Conference on Machine Learning. Lille, France: ACM, 2015: 448-456.
9	HOWARD A G, ZHU M, CHEN B, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications[EB/OL]. (2017-04-17)[2022-08-05]. .
10	ZHANG X, ZHOU X, LIN M, et al. Shufflenet: An extremely efficient convolutional neural network for mobile devices[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA: IEEE, 2018: 6848-6856.
11	GOMES R, ROZARIO P, ADHIKARI N. Deep learning optimization in remote sensing image segmentation using dilated convolutions and ShuffleNet[C]//2021 IEEE International Conference on Electro Information Technology(EIT). Mt Pleasant, MI, USA: IEEE, 2021: 244-249.
12	TAN M, LE Q. Efficientnet: Rethinking model scaling for convolutional neural networks[C]//International Conference on Machine Learning. Long Beach, California, USA: ACM, 2019: 6105-6114.
13	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA: IEEE, 2018: 7132-7141.
14	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. Yolov4: Optimal speed and accuracy of object detection[EB/OL]. (2020-04-23)[2022-08-05]. .
15	SZEGEDY C, IOFFE S, VANHOUCKE V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning[C]//Thirty-first AAAI Conference on Artificial Intelligence. San Francisco, USA: AAAI Press, 2017: 4278-4284.
16	KAISER L, GOMEZ A N, CHOLLET F. Depthwise separable convolutions for neural machine translation[EB/OL]. (2017-06-09)[2022-08-05]. .
17	伍邦谷, 张苏林, 石红, 朱鹏飞, 王旗龙, 胡清华. 基于多分支结构的不确定性局部通道注意力机制[J]. 电子学报, 2022, 50(2): 374-382.
	WU Bang-gu, ZHANG Su-lin, SHI Hong, ZHU Peng-fei, WANG Qi-long, HU Qing-hua. Multi-branch structure based local channel attention with uncertainty[J]. Acta Electronica Sinica, 2022, 50(2): 374-382. (in Chinese)
18	TAFAZZOLI F, FRIGUI H, NISHIYAMA K. A large and diverse dataset for improved vehicle make and model recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 874-881.
19	TAN M, LE Q. Efficientnetv2: Smaller models and faster training[EB/OL]. (2021-04-01)[2022-08-05]. .
20	SANDLER M, HOWARD A, ZHU M, et al. Mobilenetv2: Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA: IEEE, 2018: 4510-4520.
21	孟琭, 徐磊, 郭嘉阳. 一种基于改进的MobileNetV2网络语义分割算法[J]. 电子学报, 2020, 48(9): 1769-1776.
	MENG Lu, XU Lei, GUO Jia-yang. Semantic segmentation algorithm based on improved MobileNetV2[J]. Acta Electronica Sinica, 2020, 48(9): 1769-1776. (in Chinese)
22	MA N, ZHANG X, ZHENG H T, et al. Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]//European Conference on Computer Vision. Munich, Germany: Springer, 2018: 122-138.

数据集	类别	总数/张
常见花卉	5	3670
牛津花卉	17	2391
昆虫	3	5969
车辆^[18]	5	4028
CIFAR-10	10	60000

数据集	类别	总数/张
常见花卉	5	3670
牛津花卉	17	2391
昆虫	3	5969
车辆^[18]	5	4028
CIFAR-10	10	60000

网络	准确率/%	参数量/MB	体积/MB	训练时间/s
ResNet50	85.1	25.55	90.02	2502
ResNet50(Mish函数)	87.4	25.55	90.02	2592
ResNet50(DWConv(3×3))	84.9	14.27	46.97	2344
ResNet50(DWConv(3×3)+Maxpool)	84.1	14.27	46.98	2381
ResNet50(conv(5×5))	87.1	45.67	166.77	2728
ResNet50(PP)	89.7	25.55	90.02	2497
RemulbNet	89.0	7.50	20.98	1495
RemulbNet(DWConv+Maxpool)	86.4	7.50	20.98	1433
RemulbNet(DWConv+Avgpool)	87.9	7.50	20.98	1397

网络	准确率/%	参数量/MB	体积/MB	训练时间/s
ResNet50	85.1	25.55	90.02	2502
ResNet50(Mish函数)	87.4	25.55	90.02	2592
ResNet50(DWConv(3×3))	84.9	14.27	46.97	2344
ResNet50(DWConv(3×3)+Maxpool)	84.1	14.27	46.98	2381
ResNet50(conv(5×5))	87.1	45.67	166.77	2728
ResNet50(PP)	89.7	25.55	90.02	2497
RemulbNet	89.0	7.50	20.98	1495
RemulbNet(DWConv+Maxpool)	86.4	7.50	20.98	1433
RemulbNet(DWConv+Avgpool)	87.9	7.50	20.98	1397

网络	准确率/%	参数量 /MB	体积/MB	训练时间 /s
RemulbNet(PP)	91.6	7.50	20.98	1536
RemulbNet	89.0	7.50	20.98	1495
ResNet50	85.1	25.56	90.02	2502
MobileNet v2	81.6	3.51	8.74	2089
ShuffleNet v2	81.1	2.28	4.97	2284
EfficientNet v2	86.6	21.46	77.85	4881