基于联合剪枝深度模型压缩的种子分选方法研究

doi:10.15933/j.cnki.1004-3268.2022.01.020

摘要/Abstract

摘要： 现有基于深度学习的种子分级分选方法已取得了理想的识别效果，但由于分选方法的性能依赖于模型的宽度和深度，导致模型参数量显著增加，使其难以部署在实际应用中资源受限的边缘设备上。为此，提出了一种通道和卷积层联合剪枝的深度模型压缩方法，在通道剪枝阶段，将稀疏正则化训练的BN层参数作为衡量通道重要性的指标，在不损失模型精度的前提下，实现最大限度的通道剪枝压缩。然后提出基于线性探针的层剪枝方法，在压缩模型的同时减少内存访问，从而提升模型推理速度。最后采用知识蒸馏技术对剪枝网络进行知识迁移，补偿网络因剪枝而造成的精度损失。结果表明，在红芸豆和玉米种子数据集上所提出的方法使模型计算量减少86.55%和91.55%情况下，分别实现了实际推理速度2.1倍和2.8倍的提升，且仍保持较好的识别准确度（97.38%和96.56%），为模型在实际种子分选系统的部署提供技术支撑。

关键词: 种子分级分选, 计算机视觉, 深度学习, 模型压缩, 层剪枝, 知识蒸馏

Abstract: The crop seed sorting methods based on deep learning have achieved ideal recognition accuracy. However，their performance depends on the width and depth of the CNN（Convolutional neural network）model，which leads to a significant improvement in the number of model parameters，so it is difficult to deploy the larger‐scale model on the resource‐constrained edge devices. To address this issue，this paper proposes a depth model compression method based on joint pruning of channels and convolution layers.In the stage of channel pruning，the parameters of the BN layer trained by sparse regularization were used as the index to measure the importance of channels，to achieve the maximum channel pruning compression on the premise of keeping the accuracy of the model. Then，a layer pruning method based on a linear probe was proposed to compress the model and reduce memory access，to improve the reasoning speed of the model. Finally，the knowledge distillation technology was used in this paper to transfer the knowledge of pruning networks，thereby further compensating for the precision loss caused by pruning.The experimental results showed that，as far as the datasets of red kidney beans and maize seeds were concerned，the proposed method could improve the actual reasoning speed by 2.1 and2.8 times，while reducing the model calculation by 86. 55% and 91.55%. In the meantime，it still maintained good recognition accuracy（97.38% and 96.56%）.All these provide technical support for the deployment of the model in the actual seed sorting system.

Key words: Seed classification, Computer vision, Deep learning, Model compression, Layer pruning, Knowledge distillation

中图分类号:

S126

董燕, 李环宇, 李卫杰, 李春雷, 刘洲峰. 基于联合剪枝深度模型压缩的种子分选方法研究[J]. 河南农业科学, 2022, 51(1): 162-170.

DONG Yan, LI Huanyu, LI Weijie, LI Chunlei, LIU Zhoufeng. Research on Depth Model Compression Method Based on Joint Pruning for Seed Sorting[J]. Journal of Henan Agricultural Sciences, 2022, 51(1): 162-170.

图/表 10

参考文献

［1］ALTUNTAS Y，OCAMAZ A F，ENGIZ R，et al.Classification of haploid and diploid maize seeds by using image processing techniques and support vector machines［C］//26th signal processing and communications applications conference. Cesme：IEEE，2018：1‐4.

［2］祝保林，张彦波，张欣羽，等.基于机器视觉的小桐子种子识别算法研究［J］.种子，2020，39（10）：158‐159，164.

ZHU B L，ZHANG Y B，ZHANG X Y，et al.Research on seed recognition algorithm of jatropha curcas seeds based on machine vision［J］.Seed，2020，39（10）：158‐159，164.

［3］LIU D，NING X，LI Z，et al.Discriminating and elimination of damaged soybean seeds based on image characteristics［J］.Journal of Stored Products Research，2014，60：67‐74.

［4］黄志强，李军.基于卷积神经网络的图像识别研究综述［J］.汽车工程师，2020（10）：11‐13，31.

HUANG Z Q，LI J.Review of research on image recognition based on convolutional neural network［J］.Auto Engineer，2020（10）：11‐13，31.

［5］范丽丽，赵宏伟，赵浩宇，等.基于深度卷积神经网络的目标检测研究综述［J］.光学精密工程，2020，28（5）：1152‐1164.

FAN L L，ZHAO H W，ZHAO H Y，et al.Survey of target detection based on deep convolutional neural networks ［J］.Optics and Precision Engineering，2020，28（5）：1152‐1164.

［6］田启川，孟颖.卷积神经网络图像语义分割技术［J］.小型微型计算机系统，2020，41（6）：1302‐1313.

TAN Q C，M Y.Image semantic segmentation based on convolutional neural network［J］.Journal of Chinese Computer Systems，2020，41（6）：1302‐1313.

［7］SIMONYAN K，ZISSERMAN A.Very deep convolutional networks for large‐scale image recognition［J/OL］.（2015‐04‐10）［2021‐05‐10］.https：// arxiv. org/abs/1409. 1556.

［8］HE K K，ZHANG X G，REN S Q，et al.Deep residual learning for image recognition［C］//29st IEEE/CVF.Conference on computer vision and pattern recognition.Las Vegas：IEEE，2016：770‐778.

［9］SZEGEDY C，WEI L，JIA Y，et al.Going deeper with convolutions ［C］//28st IEEE/CVF. Conference on computer vision and pattern recognition. Boston：IEEE，2015：1‐9.

［10］冯晓，李丹丹，王文君，等.基于轻量级卷积神经网络和迁移学习的小麦叶部病害图像识别［J］.河南农业科学，2021，50（4）：174‐180.

FENG X，LI D D，WANG W J，et al.Image recognition of wheat leaf diseases based on lightweight convolutional neural network and transfer learning［J］.Journal of Henan Agricultural Sciences，2021，50（4）：174‐180.

［11］张永超，赵录怀，卜光苹.基于卷积神经网络的花生种子筛选识别算法［J］.江西农业学报，2020，32（1）：77‐82.

ZHANG Y C，ZHAO H L，BU G P.Screening and recognition algorithm of peanut seed based on convolutional neural network［J］.Jiangxi Agricultural Journal，2020，32（1）：77‐82.

［12］朱荣胜，闫学慧，陈庆山.基于图像识别和卷积神经网络的大豆优良籽粒筛选研究［J］.大豆科学，2020，39（2）：189‐197.

ZHU R S，YAN X X，CHEN Q S.Study on the optimization of soybean seed selection based on image recognition and convolution neural network［J］.Soybean Science，2020，39（2）：189‐197.

［13］HUANG S，FAN X F，SUN L，et al.Research on classification method of maize seed defect based on machine vision［J］.Journal of Sensors，2019，2019：1‐9.

［14］HAN S，POOL J，TRAN J，et al.Learning both weights and connections for efficient neural networks［J/OL］.（2015‐10‐01）［2021‐05‐10］.https：//arxiv.org/abs/1506. 02626.

［15］LIN M，JI R，WANG Y，et al.HRank：Filter pruning using high‐rank feature map［C］//33st IEEE/CVF.Conference on computer vision and pattern recognition Seattle：IEEE，2020：1529‐1538.

［16］ZHUANG L，LI J，SHEN Z，et al.Learning efficient convolutional networks through network slimming［C］//IEEE.2017 IEEE international conference on computer vision. Venice：IEEE，2017：2736‐2744.

［17］CORREIA A L，LIE M，SCHWARTZ W R.Discriminative layer pruning for convolutional neural network［J］.IEEE Journal of Selected Topics in Signal Processing，2020，14（4）：828‐837.

［18］ELKERDAWY S，ELHOUSHI M，SINGH A，et al.To filter prune，or to layer prune，that is the question［J/OL］.（2020‐11‐08）［2021‐05‐10］.https：//arxiv.org/abs/2007. 05667.

［19］HINTON G，VINYALS O，DEAN J.Distilling the knowledge in a neural network［J/OL］.（2015‐03‐09）［2021‐05‐10］.https：//arxiv. org/abs/1503. 02531.

［20］ALAIN G，BENGIO Y.Understanding intermediate layers using linear classifier probes［J/OL］.（2018‐11‐22）［2021‐05‐10］.https：//arxiv.org/abs/1610.016441.

［21］DING X，ZHANG X，MA N，et al.Repvgg：making vgg‐style convnets great again［C］//34st IEEE/CVF.Conference on computer vision and pattern recognition.Online：IEEE，2021：13733‐13742.

［22］ZHANG T，YE S，ZHANG K，et al.A systematic dnn weight pruning framework using alternating direction method of multipliers［C］//CVF.2018 European conference on computer vision. Munich：Springer，2018：184‐199.

［23］ALTUNTAŞ Y，COMERT R，KOCAMAZ A F.Identification of haploid and diploid maize seeds using convolutional neural networks and a transfer learning approach［J］.Computers and Electronics in Agriculture，2019，163：104874.

[1]	于雁南, 莫泳彬, 严继池, 熊春林, 窦世卿, 杨荣峰. 基于改进ShuffleNet V2 的柑橘病害识别研究[J]. 河南农业科学, 2024, 53(1): 142-151.
[2]	钟正扬, 云利军, 杨璇玺, 陈载清. 基于改进YOLOX 的自然环境下核桃识别算法研究[J]. 河南农业科学, 2024, 53(1): 152-161.
[3]	李敏, 郭雷风, 王瑞利. 无人机遥感在马铃薯田间表型研究中的应用[J]. 河南农业科学, 2023, 52(9): 24-32.
[4]	李伟豪, 詹炜, 周婉, 韩涛, 王佩文, 刘虎, 熊梦园, 孙泳, . 轻量型Yolov7-TSA 网络在茶叶病害检测识别中的研究与应用[J]. 河南农业科学, 2023, 52(5): 162-169.
[5]	陈聪, 于啸, 宫琪. 基于改进残差网络的苹果叶片病害识别研究[J]. 河南农业科学, 2023, 52(4): 152-161.
[6]	张红涛, 罗一铭, 谭联, 杨加蓬, 王宇. 基于迁移学习和残差网络的谷子病害识别研究[J]. 河南农业科学, 2023, 52(12): 162-171.
[7]	臧贺藏, 王从胜, 赵巧丽, 赵晴, 张杰, 李国强, 郑国清. 基于深度学习的小麦倒伏自动分类方法研究[J]. 河南农业科学, 2023, 52(11): 167-173.
[8]	杨文庆, 刘天霞, 唐兴萍, 徐国富, 马喆, 杨贺凯, 吴文斗. 智慧农业背景下的植物表型组学研究进展[J]. 河南农业科学, 2022, 51(7): 1-12.
[9]	刁智华, 闫娇楠, 赵素娜, 贺振东. 基于图像处理的作物行识别算法研究进展[J]. 河南农业科学, 2022, 51(3): 12-19.
[10]	冯晓, 李丹丹, 王文君, 郑国清, 刘海礁, 孙永胜, 梁山, 杨莹, 臧贺藏, 张辉. 基于轻量级卷积神经网络和迁移学习的小麦叶部病害图像识别[J]. 河南农业科学, 2021, 50(4): 174-180.
[11]	郑二功，田迎芳，陈涛. 基于深度学习的无人机影像玉米倒伏区域提取[J]. 河南农业科学, 2018, 47(8): 155-160.