Lightweight Beverage Recognition Network Based on GhostNet Residual Structure

doi:10.19678/j.issn.1000-3428.0059966

Abstract

Abstract: To alleviate the problem that the Yolov4-Tiny algorithm cannot be deployed on a platform with limited resources due to its large number of parameters and heavy computational requirements, a lightweight target detection network based on GhostNet residue structure is proposed in this study.The proposed network adopts the GhostNet structure to divide ordinary convolution into two steps, that is, a small number of convolution kernels are used to generate one part of the feature graph, and the other part of the generated feature graph is obtained through simple calculation, and the two groups of feature graphs are spliced, which reduces the required resource load and number of parameters.This study indicates that the normalization layer of Yolov4-Tiny algorithm reduces the model size leveraging the residue structure built by GhostNet to only 2.18 MB, which is 90% smaller than that of the original Yolov4-Tiny algorithm.The average image processing speed of Yolov4-Ghostnet algorithm is 24% faster than that of Yolov4-Tiny algorithm with respect to GPU acceleration and 56% faster regarding CPU processing.The experimental results indicate that the algorithm achieves a mean Average Precision(mAP) value of 79.43% on the beverage test set.Compared to the Yolov4-Tiny algorithm without loss of accuracy, the proposed algorithm significantly reduces the amount of network computational load and its number of parameters, speeds up the reasoning speed, and is more suitable to be deployed in embedded devices with insufficient resource computing power.

Key words: deep learning, Convolutional Neural Network(CNN), YOLOv4-Tiny algorithm, residual structure, lightweight, object detection

摘要： YOLOv4-Tiny目标检测网络算法存在参数多和计算量大等问题，无法部署在资源有限的平台上。提出一种基于GhostNet残差结构的主干轻量级目标检测网络算法YOLO-GhostNet。该算法采用GhostNet结构将普通卷积分成两步，即使用较少的卷积核生成一部分特征图，对生成的特征图通过简单计算获得另一部分特征图，并将两组特征图进行拼接，以减少计算所需资源与参数量。通过GhostNet构建残差结构的YOLO-GhostNet算法在经过批量归一化层优化后模型尺寸只有2.18 MB，较YOLOv4-Tiny算法模型尺寸减小90%。YOLO-GhostNet算法在GPU加速环境下平均处理图片速度比YOLOv4-Tiny算法提高24%，CPU处理速度比YOLOv4-Tiny加快56%。实验结果表明，该算法在饮料测试集中的平均精确度均值达到79.43%，相比YOLOv4-Tiny算法，其在精度无损失情况下能够大幅降低网络计算量和参数量，同时加快推理速度，更适合部署于资源算力不足的嵌入式设备。

关键词: 深度学习, 卷积神经网络, YOLOv4-Tiny算法, 残差结构, 轻量化, 目标检测

CLC Number:

TP183

CAO Yuanjie, GAO Yuxiang. Lightweight Beverage Recognition Network Based on GhostNet Residual Structure[J]. Computer Engineering, 2022, 48(3): 310-314.

曹远杰, 高瑜翔. 基于GhostNet残差结构的轻量化饮料识别网络[J]. 计算机工程, 2022, 48(3): 310-314.

/ / Recommend / Download Citations

URL: http://www.ecice06.com/EN/10.19678/j.issn.1000-3428.0059966

http://www.ecice06.com/EN/Y2022/V48/I3/310

Figures/Tables 7

References

[1] SIMONYAN K, ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[EB/OL].[2020-10-10].http://www.arxiv.org/pdf/1409.1556.pdf.
[2] SZEGEDY C, LIU W, JIA Y, et al.Going deeper with convolutions[C]//Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2014:1-9.
[3] HUNG G, LIU Z, MAATEN L V D, et al.Densely connected convolutional networks[C]//Proceedings of IEEE CVPRʼ17.Washington D.C., USA:IEEE Press, 2017:235-246.
[4] MOLCHANOV P, MALLYA A, TYREE S, et al.Importance estimation for neural network pruning[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2019:11264-11272.
[5] ZHOU S, WU Y, NI Z, et al.DoReFa-Net:training low bitwidth convolutional neural networks with low bitwidth gradients[EB/OL].[2020-10-10]. https://arxiv.org/abs/1606.06160.
[6] HINTON G, VINYALS O, DEAN J.Distilling the knowledge in a neural network[J].Computerence, 2015, 14(7):38-39.
[7] KIM D H.Evaluation of COCO validation 2017 dataset with YOLOv3[J].Evaluation, 2019, 6(7):10356-10360.
[8] REDMON J, FARHADI A.YOLOv3:an incremental improvement[EB/OL].[2020-10-10].http://arxiv.org/pdf/1804.02767.
[9] 邵伟平, 王兴, 曹昭睿, 等.基于MobileNet与YOLOv3的轻量化卷积神经网络设计[J].计算机应用, 2020, 40(z1):8-13. SHAO W P, WANG X, CAO Z R, et al.Design of lightweight convolutional neural network based on MobileNet and YOLOv3[J].Journal of Computer Applications, 2020, 40(z1):8-13.(in Chinese)
[10] 刘万军, 高明月, 曲海成, 等.基于反残差结构的轻量级多目标检测网络[J].激光与光电子学进展, 2019, 56(22):65-73. LIU W J, GAO M Y, QU H C.Light-weight multi-object detection network based on inverted residual structure[J].Laser and Optoelectronics Progress, 2019, 56(22):65-73.(in Chinese)
[11] CHOLLET F.Xception:deep learning with depth wise separable convolutions[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2017:1251-1258.
[12] IANDOLA F N, HAN S, MOSKEWICZ M W, et al.SqueezeNet:AlexNet-level accuracy with 50x fewer parameters and <0.5 MB model size[EB/OL].[2020-10-10].http://arxiv.org/pdf/1602.07360.
[13] HOWARD A G, ZHU M, CHEN B, et al.MobileNets:efficient convolutional neural networks for mobile vision applications[EB/OL].[2020-10-10].http://arxiv.org/pdf/1704.04861.
[14] HOWARD A, SANDLER M, CHU G, et al.Searching for MobileNetV3[EB/OL].[2020-10-10].http://arxiv.org/abs/1905.02244v3.
[15] TAN M, LE Q V.Efficientnet:rethinking model scaling for convolutional neural networks[EB/OL].[2020-10-10].http://arxiv.org/abs/1905.11946.
[16] HE K, ZHANG X, REN S, et al.Deep residual learning for image recognition[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2016:145-158.
[17] HAN K, WANG Y, TIAN Q, et al.GhostNet:more features from cheap operations[C]//Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Seattle, USA:IEEE Press, 2020:1577-1586.
[18] REDMON J, DIVVALA S, GIRSHICK R, et al.You only look once:unified, real-time object detection[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2016:779-788.
[19] REDMON J, FARHADI A.YOLO9000:better, faster, stronger[C]//Proceedings of IEEE Conference on Computer Vision & Pattern Recognition.Washington D.C., USA:IEEE Press, 2017:6517-6525.
[20] BOCHKOVSKIY A, WANG C Y, LIAO H Y M.YOLOv4:optimal speed and accuracy of object detection[EB/OL].[2020-10-10].http://arxiv.org/abs/2004.10934.
[21] XU Z F, JIA R S, LIU Y B, et al.Fast method of detecting tomatoes in a complex scene for picking robots[J].IEEE Access, 2020, 7:11123-111135.
[22] 管军霖, 智鑫.基于YOLOv4卷积神经网络的口罩佩戴检测方法[J].现代信息科技, 2020, 4(11):9-12. GUAN J L, ZHI X.Mask wearing detection method based on YOLOv4 convolutional neural network[J].Modern Information Technology, 2020, 4(11):9-12.(in Chinese)

Please choose a citation manager

Content to export