基于改进EfficientNetV2模型的苹果叶片病害识别

doi:10.16178/j.issn.0528-9017.20240918

摘要/Abstract

摘要：

苹果叶片病害会严重影响苹果的产能和品质,精准识别病害对于有效防治和减少损失具有重要意义。苹果叶片病斑大小不一且在叶片上的分布位置复杂多样,同时还受到自然环境中的复杂背景干扰,导致叶片病害难以识别、准确率低等问题。基于此,本文提出了一个基于注意力机制改进EfficientNetV2的苹果叶片病害识别模型(Efficient-BEANet)。首先,本研究提出了一种高速双通道归一化(BCET)方法,建立了批归一化(batch normalization, BN)和层归一化(layer normalization, LN)方法之间的联系,使模型能够融合多个尺度的特征信息,进而改善小病斑识别困难的问题,同时将该模型与增强线性变换相结合,以加快模型的收敛速度。此外,通过设计多尺度共享注意力机制(ESCA),建立空间和通道2个维度特征之间的相关性,通过自动学习病害的特征,增强重要特征表示,对出现病斑的区域进行重点关注。最后,通过提取多层次共享的通道特征,自适应地为重要特征分配更大的权重,并抑制与病害无关的信息,以期有效改善因病斑不规则的空间分布以及复杂背景干扰而造成的误检漏检问题。结果表明,改进后的模型在验证集上的准确率达到93.32%,相较于原模型提升3.49百分点,适用于对苹果叶片病害的识别。

关键词: 苹果叶片病害, 归一化, 注意力机制

Abstract:

Apple leaf diseases can seriously affect the production capacity and quality of apples. Accurate identification of diseases is of great significance for effective prevention and control and reducing losses. The lesions of apple leaf diseases vary in size and are distributed in complex and diverse positions on the leaves. Moreover, they are disturbed by the complex background in the natural environment, which leads to problems such as difficulty in identifying leaf diseases and low accuracy. Based on this, this paper proposed an apple leaf disease recognition model (Efficient-BEANet) that improved EfficientNetV2 based on the attention mechanism. Firstly, this study proposed a high-speed dual-channel normalization (BCET) method and established the connection between batch normalization (BN) and layer normalization (LN) methods, enabling the model to integrate feature information of multiple scales and thereby improve the problem of difficult identification of small lesions. Meanwhile, this model was combined with the enhanced linear transformation to accelerate the convergence speed of the model. In addition, by designing a multi-scale shared attention mechanism (ESCA), the correlation between the features of the two dimensions of space and channel was established. Through automatic learning of the features of diseases, the representation of important features was enhanced, and the areas where disease spots appear were given special attention. Finally, by extracting multi-level shared channel features, greater weights were adaptively assigned to important features, and information unrelated to the disease was suppressed, with the aim of effectively improving the problems of false detection and missed detection caused by the irregular spatial distribution of disease spots and the interference of complex backgrounds. The experimental results showed that the accuracy of the improved model on the validation set reached 93.32%, which was increased by 3.49 percentage points compared with the original model, and can be used for the effective identification of apple leaf diseases.

Key words: apple leaf disease, normalization, attention mechanism

中图分类号:

S436.611

张鹏, 刘朔, 欧阳宇, 李萌民. 基于改进EfficientNetV2模型的苹果叶片病害识别[J]. 浙江农业科学, 2026, 67(1): 115-124.

ZHANG Peng, LIU Shuo, OUYANG Yu, LI Mengmin. Apple leaf disease recognition based on improved EfficientNetV2 model[J]. Journal of Zhejiang Agricultural Sciences, 2026, 67(1): 115-124.

图/表 16

参考文献 21

[1]	陈学森, 王楠, 彭福田, 等. 中国重要落叶果树果实品质和熟期育种研究进展[J]. 园艺学报, 2024, 51(1): 8-26.
	CHEN X S, WANG N, PENG F T, et al. Advances in quality and maturity breeding of important deciduous fruit trees in China[J]. Acta Horticulturae Sinica, 2024, 51(1): 8-26.
[2]	王浩宇, 崔艳荣, 胡玉荣, 等. 基于改进EfficientNetV2的苹果叶片病害识别模型[J]. 山东农业科学, 2024, 56(9): 124-132.
	WANG H Y, CUI Y R, HU Y R, et al. Apple leaf disease recognition model based on improved EfficientNetV2[J]. Shandong Agricultural Sciences, 2024, 56(9): 124-132.
[3]	王瑞鹏, 陈锋军, 朱学岩, 等. 采用改进的EfficientNet识别苹果叶片病害[J]. 农业工程学报, 2023, 39(18): 201-210.
	WANG R P, CHEN F J, ZHU X Y, et al. Identifying apple leaf diseases using improved EfficientNet[J]. Transactions of the Chinese Society of Agricultural Engineering, 2023, 39(18): 201-210.
[4]	刘小玲, 崔艳荣. 基于改进轻量化网络MobileNeXt的苹果叶片病害识别方法[J]. 江苏农业科学, 2023, 51(10): 185-193.
	LIU X L, CUI Y R. An improved method for apple leaf disease identification based on lightweight network MobileNeXt[J]. Jiangsu Agricultural Sciences, 2023, 51(10): 185-193.
[5]	HOU Q B, ZHOU D Q, FENG J S. Coordinate attention for efficient mobile network design[C]// 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 20-25, 2021. Nashville, TN, USA. IEEE, 2021: 13708-13717.
[6]	CHEN S R. CNN combined with data augmentation for face recognition on small dataset[J]. Journal of Physics: Conference Series, 2023, 2634(1): 012040.
[7]	KHALED A. BCN: batch channel normalization for image classification[C]// International Conference on Pattern Recognition. Springer, Cham, 2025.
[8]	XU Y, XIE L, XIE C, et al. BNET: batch normalization with enhanced linear transformation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(7):8.
[9]	MNIH V, HEESS N, GRAVES A. Recurrent models of visual attention[J]. Advances in Neural Information Processing Systems, 2014, 27.
[10]	SI Y, XU H, ZHU X, et al. SCSA: exploring the synergistic effects between spatial and channel attention[J]. Neurocomputing, 2025, 634.
[11]	HUANG H J, CHEN Z G, ZOU Y, et al. Channel prior convolutional attention for medical image segmentation[J]. Computers in Biology and Medicine, 2024, 178: 108784.
[12]	WANG Q L, WU B G, ZHU P F, et al. ECA-net: efficient channel attention for deep convolutional neural networks[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 13-19, 2020. Seattle, WA, USA. IEEE, 2020: 11531-11539.
[13]	HU J, SHEN L, ALBANIE S, et al. Gather-excite: exploiting feature context in convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2018, 31.
[14]	LIU Y, SHAO Z, TENG Y, et al. NAM: Normalization-based attention module[EB/OL]. ( 2021-11-24)[2024-08-20]. https://arxiv.org/abs/2111.12419.
[15]	OUYANG D L, HE S, ZHANG G Z, et al. Efficient multi-scale attention module with cross-spatial learning[C]// ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). June 4-10, 2023, Rhodes Island, Greece. IEEE, 2023: 1-5.
[16]	WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]// Computer Vision-ECCV 2018. Cham: Springer International Publishing, 2018: 3-19.
[17]	SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]// 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 7-12, 2015, Boston, MA, USA. IEEE, 2015: 1-9.
[18]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016, Las Vegas, NV, USA. IEEE, 2016: 770-778.
[19]	ZHANG X Y, ZHOU X Y, LIN M X, et al. ShuffleNet: an extremely efficient convolutional neural network for mobile devices[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. June 18-23, 2018, Salt Lake City, UT, USA. IEEE, 2018: 6848-6856.
[20]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). July 21-26, 2017, Honolulu, HI, USA. IEEE, 2017: 2261-2269.
[21]	XU J, PAN Y, PAN X L, et al. RegNet: self-regulated network for image classification[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023, 34(11): 9562-9567.

编号	名称	扩展因子	卷积核(步长)	通道数	层数
—	3×3卷积	—	3×3(2)	24	1
stage 1	EBF-MBConv	1	3×3(1)	24	2
stage 2	EBF-MBConv	4	3×3(2)	48	4
stage 3	EBF-MBConv	4	3×3(2)	64	4
stage 4	EB-MBConv	4	3×3(2)	128	6
stage 5	EB-MBConv	6	3×3(1)	160	9
stage 6	EB-MBConv	6	3×3(2)	256	15
—	1×1卷积层&池化层&全连接层	—	—	1 280	1

编号	名称	扩展因子	卷积核(步长)	通道数	层数
—	3×3卷积	—	3×3(2)	24	1
stage 1	EBF-MBConv	1	3×3(1)	24	2
stage 2	EBF-MBConv	4	3×3(2)	48	4
stage 3	EBF-MBConv	4	3×3(2)	64	4
stage 4	EB-MBConv	4	3×3(2)	128	6
stage 5	EB-MBConv	6	3×3(1)	160	9
stage 6	EB-MBConv	6	3×3(2)	256	15
—	1×1卷积层&池化层&全连接层	—	—	1 280	1

归一化方法	精准度	召回率	F₁值	准确率
批归一化(BN)	89.95	89.73	89.76	89.83
层归一化(LN)	87.69	87.10	87.26	87.28
同步批量归一化(SN)	89.59	88.38	88.68	88.55
实例归一化(IN)	89.55	89.05	89.13	89.18
局部响应归一化(LRN)	89.94	89.18	89.40	89.34
增强线性变换的批量归一化 (BNET)	88.02	87.12	87.36	87.28
批量通道归一化(BCN)	89.10	87.88	88.17	88.07
高速双通道归一化(BCET)	92.36	92.42	92.33	92.53

归一化方法	精准度	召回率	F₁值	准确率
批归一化(BN)	89.95	89.73	89.76	89.83
层归一化(LN)	87.69	87.10	87.26	87.28
同步批量归一化(SN)	89.59	88.38	88.68	88.55
实例归一化(IN)	89.55	89.05	89.13	89.18
局部响应归一化(LRN)	89.94	89.18	89.40	89.34
增强线性变换的批量归一化 (BNET)	88.02	87.12	87.36	87.28
批量通道归一化(BCN)	89.10	87.88	88.17	88.07
高速双通道归一化(BCET)	92.36	92.42	92.33	92.53

类型	名称	参数量/10⁶	精准度/%	召回率/%	F₁值/%	准确率/%
通道注意力	通道注意力机制(SE)	19.25	89.95	89.73	89.76	89.83
	高效通道注意力机制(ECA)	19.25	89.49	88.91	89.02	89.03
	聚集-激发注意力机制(GE)	24.51	88.39	88.20	88.61	89.76
归一化注意力	非局部注意力机制(NAM)	19.32	88.50	88.09	88.19	88.24
无参数注意力	简单注意力机制(SimAM)	19.25	76.74	74.23	74.27	74.09
空间与通道注意力	高效多尺度注意力机制(EMA)	19.66	88.31	87.49	87.66	87.59
	卷积块注意力机制(CBAM)	24.42	86.39	86.17	86.14	86.33
	协调注意力机制(CA)	26.99	88.94	88.59	88.71	88.71
	多尺度共享注意力机制(ESCA)	24.96	92.13	92.13	92.14	92.26