FLPNet：一种基于频空双域解耦的轻量级实时苹果叶片病害检测模型

doi:10.12133/j.smartag.SA202605024

摘要/Abstract

摘要：

【目的/意义】 在复杂成像环境中，动态光照变化和叶片遮挡会降低病斑特征的可见性，使现有模型难以准确识别早期微小病斑。为解决自然环境中显著光照干扰所带来的苹果叶片病害检测难题，提出了一种改进的RT-DETR（Real-Time Detection Transformer）模型，命名为频域-光照感知网络（Frequency-Lighting-Aware Perception Network, FLPNet），用于复杂成像条件下的病害检测。 【方法】 为克服传统卷积神经网络在空间域特征提取方面的局限性，FLPNet引入了频域解耦与增强模块（Frequency-Domain Decoupling and Enhancement Module, FDEM）。通过利用小波变换与傅里叶分析的互补特性，FDEM有效解耦并重构了受遮挡病斑的细粒度纹理特征。此外，引入了物理驱动Retinex层，该层作为一种基于Retinex理论的特征增强模块，用于减轻非受控光照的影响。同时，还嵌入了低分辨率全局注意力模块，以在保留全局上下文信息的同时减少计算冗余。 【结果和讨论】 FLPNet的mAP@0.50达到85.7%，较基线模型提高了5.0个百分点，同时参数量减少了3.3 M，计算量降低了9.1 GFLOPs。与YOLOv12n相比，FLPNet的mAP@0.75提高了4.1个百分点。实验结果表明，FLPNet在整体检测性能上优于YOLOv12n和DEIM等先进模型。 【结论】 FLPNet在显著降低计算复杂度的同时，有效提升了复杂自然成像条件下苹果叶片病害的检测性能。该模型在检测性能与计算效率之间取得了较好的平衡，适合部署于资源受限的农业边缘设备。这些特性可为苹果叶片病害检测的实际应用提供重要的理论支持和技术指导。

关键词: 苹果叶片病害, RT-DETR, 注意力机制, Retinex理论, 轻量化

Abstract:

[Objective] In complex imaging environments, dynamic lighting variations and foliage occlusion can degrade the visibility of disease spot features, making the accurate identification of early-stage micro-lesions difficult for existing models. To address the challenges associated with apple leaf disease detection in natural environments characterized by significant lighting interference, an enhanced real-time detection transformer (RT-DETR) model, termed the frequency-lighting-aware perception network (FLPNet), is proposed for disease detection under complex imaging conditions. [Methods] FLPNet was developed to overcome the limitations of traditional convolutional neural networks in spatial-domain feature extraction by incorporating a frequency-domain decoupling and enhancement module (FDEM). By leveraging the complementary properties of wavelet transform and Fourier analysis, FDEM effectively decoupled and reconstructed fine-grained texture features of occluded disease spots. Furthermore, a physics-driven Retinex layer, designed as a feature enhancement module based on Retinex theory, was introduced to mitigate the effects of uncontrolled illumination. A low-resolution global attention module was also incorporated to reduce computational redundancy while preserving global contextual information. [Results and Discussions] FLPNet achieved an mAP@0.50 of 85.7%, representing an improvement of 5.0 percentage points, while reducing the number of parameters by 3.3 M and the computational cost by 9.1 GFLOPs. Compared with YOLOv12n, FLPNet achieved a 4.1 percentage points improvement in mAP@0.75. The results showed that FLPNet outperformed state-of-the-art models, including YOLOv12n and DEIM, in overall detection performance. The proposed modules significantly reduced computational complexity while enhancing disease detection capability under complex imaging conditions. [Conclusions] FLPNet improves apple leaf disease detection under complex natural imaging conditions while substantially reducing computational complexity. The model achieves a favorable balance between detection performance and computational efficiency, indicating potential for deployment on resource-constrained agricultural edge devices. These characteristics provide essential theoretical support and technical guidance for the practical application of apple leaf disease detection.

Key words: apple leaf diseases, RT-DETR, attention mechanism, Retinex theory, lightweight

中图分类号:

TP183

宋祥硕, 王钧. FLPNet：一种基于频空双域解耦的轻量级实时苹果叶片病害检测模型[J]. 智慧农业(中英文), 2026, 8(5): 219-237.

SONG Xiangshuo, WANG Jun. FLPNet: A Lightweight Real-Time Apple Leaf Disease Detection Model Based on Frequency-Spatial Dual-Domain Decoupling[J]. Smart Agriculture, 2026, 8(5): 219-237.

图/表 21

Disease type	Scientific name	Symptoms	Period of high incidence
Alternaria leaf spot^[15]	Alternaria spp	Dark brown spots appear on the leaves, evolving into circular or irregular lesions characterized by grayish-brown centers and occasionally concentric rings. The necrotic tissue leads to premature wilting and dropping of the leaves	It mainly occurs between 70 and 110 days after flowering, with infection typically starting around 20 days after flowering
Brown spot^[16]	Alternaria alternata f. sp. mali	Dark brown spots expand into larger lesions, accompanied by a rapid yellowing of the surrounding tissue, forming a distinct "halo". This process leads to extensive premature leaf defoliation	From summer to early September, conditions favoring early defoliation disease are prevalent due to the hot and rainy climate. The peak period for this disease extends until early September
Frogeye leaf spot^[17]	Botryosphaeria dothidea	Purple-brown specks develop into circular spots characterized by light centers and dark margins, resembling a "frog's eye". The necrotic center may detach, potentially leading to leaf drop	Its peak period usually occurs from late spring to early summer
Mosaic^[18]	Apple mosaic virus	Light and dark green mosaic patterns appear on leaves, with yellow patches reducing chlorophyll. This impairs photosynthesis and weakens the tree	Symptoms can occur throughout the year, but are often more pronounced during the spring shoot growth period
Powdery mildew^[19]	Podosphaera leucotricha	White powdery patches form on the leaves, causing them to curl and distort. The affected areas become stunted, potentially leading to premature leaf loss	Spring is the peak season for the occurrence of diseases; however, the severity of these diseases varies across different regions. In some areas, diseases may persist throughout other periods of the growing season
Rust^[20]	Gymnosporangium yamadae	Yellow-orange spots appear on the leaf surface, accompanied by corresponding cup-shaped structures forming beneath. The leaves may curl, desiccate, and abscise prematurely	Spring and early summer represent the peak periods for rust disease, particularly in the Shaanxi production area, where a high incidence of rust primarily occurs during the early to middle stages of the growing season
Scab^[21]	Venturia inaequalis	Velvety, olive-green to black spots develop on the leaves, evolving into raised, scabby lesions. The surrounding tissue may exhibit yellowing, which can lead to curling and premature leaf drop	Spring is a critical period for the occurrence of apple scab, as it is closely linked to the rainfall and humidity conditions prevalent during this season

参考文献 41

[1]	ZHANG Q Q, SHI F J, ABDULLAHI N M, et al. An empirical study on spatial-temporal dynamics and influencing factors of apple production in China[J]. PLoS One, 2020, 15(10): e0240140.
[2]	KHAN A I, QUADRI S M K, BANDAY S, et al. Deep diagnosis: A real-time apple leaf disease detection system based on deep learning[J]. Computers and Electronics in Agriculture, 2022, 198: 107093.
[3]	SAMAJPATI B J, DEGADWALA S D. Hybrid approach for apple fruit diseases detection and classification using random forest classifier[C]// 2016 International Conference on Communication and Signal Processing (ICCSP). Piscataway, New Jersey, USA: IEEE, 2016: 1015-1019.
[4]	AZGOMI H, HAREDASHT F R, SAFARI MOTLAGH M R. Diagnosis of some apple fruit diseases by using image processing and artificial neural network[J]. Food Control, 2023, 145: 109484.
[5]	ALI K M. Detection and classification of plant diseases using image processing and multiclass support vector machine[J]. International Journal of Computer Trends and Technology, 2020, 68(4): 5-11.
[6]	LI T, ZHANG L Y, LIN J C. Precision agriculture with YOLO-Leaf: Advanced methods for detecting apple leaf diseases[J]. Frontiers in Plant Science, 2024, 15: 1452502.
[7]	SUN Z K, FENG Z M, CHEN Z M. Highly accurate and lightweight detection model of apple leaf diseases based on YOLO[J]. Agronomy, 2024, 14(6): 1331.
[8]	LI L L, ZHANG S J, WANG B. Apple leaf disease identification with a small and imbalanced dataset based on lightweight convolutional networks[J]. Sensors, 2021, 22(1): 173.
[9]	WANG S, LIU M Y, DONG S C, et al. ERL-RTDETR: A lightweight transformer-based framework for high-accuracy apple disease detection in precision agriculture[J]. Concurrency and Computation: Practice and Experience, 2025, 37(23/24): e70276.
[10]	FU Z L, YIN L F, CUI C, et al. A lightweight MHDI-DETR model for detecting grape leaf diseases[J]. Frontiers in Plant Science, 2024, 15: 1499911.
[11]	JI J T, YANG Y Q, MA H, et al. A lightweight detection method for common peach tree diseases based on the FHLE-RTDETR model[J]. Smart Agricultural Technology, 2025, 12: 101352.
[12]	YAO J B, ZHU Z H, YUAN M Q, et al. The detection of maize leaf disease based on an improved real-time detection transformer model[J]. Symmetry, 2025, 17(6): 808.
[13]	CHEN Z Y, YANG Y H. Grape leaf disease detection and recognition method based on improved model RTDETR-DRA[C]// 2024 4th International Conference on Electronic Information Engineering and Computer (EIECT). Piscataway, New Jersey, USA: IEEE, 2024: 280-285.
[14]	HUGHES D P, SALATHE M. An open access repository of images on plant health to enable the development of mobile disease diagnostics[PP/OL]. arXiv: 1511.08060, 2016.
[15]	ZHANG Q L, XU C R, WEI H Y, et al. Two pathogenesis-related proteins interact with leucine-rich repeat proteins to promote Alternaria leaf spot resistance in apple[J]. Horticulture Research, 2021, 8: 219.
[16]	HOU Y M, ZHANG X, ZHANG N N, et al. Genera Acremonium and Sarocladium cause brown spot on bagged apple fruit in China[J]. Plant Disease, 2019, 103(8): 1889-1901.
[17]	DUTTA P, DEB L, DEBBARMA M. First report of Diplodia seriata de Not. causing black rot of apple (Malus domestica L. Borkh.) in Meghalaya, North Eastern India[J]. The Pharma Innovation, 2020, 9(6): 388-391.
[18]	GRIMOVÁ L, WINKOWSKA L, KONRADY M, et al. Apple mosaic virus[J]. Phytopathologia Mediterranea, 2016, 55(1): 1-19.
[19]	STRICKLAND D A, HODGE K T, COX K D. An examination of apple powdery mildew and the biology of Podosphaera leucotricha from past to present[J]. Plant Health Progress, 2021, 22(4): 421-432.
[20]	WANG Y, AN H, YANG Y, et al. The MpNAC72/MpERF105-MpMYB10b module regulates anthocyanin biosynthesis in Malus 'Profusion' leaves infected with Gymnosporangium yamadae [J]. The Plant Journal, 2024, 118(5): 1569-1588.
[21]	KHAJURIA Y P, AKHOON B A, KAUL S, et al. Secretomic insights into the pathophysiology of Venturia inaequalis: the causative agent of scab, a devastating apple tree disease[J]. Pathogens, 2022, 12(1): 66.
[22]	XU M L, YOON S, FUENTES A, et al. A comprehensive survey of image augmentation techniques for deep learning[J]. Pattern Recognition, 2023, 137: 109347.
[23]	LI F F, ZHANG Y H, FAN Y X. EABI-DETR: an efficient aerial small object detection network[J]. Biomimetics, 2025, 10(11): 770.
[24]	HAN K, WANG Y H, CHEN H T, et al. A survey on vision transformer[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(1): 87-110.
[25]	LI J F, HAO S, LI T S, et al. RDMA: low-light image enhancement based on retinex decomposition and multi-scale adjustment[J]. International Journal of Machine Learning and Cybernetics, 2024, 15(5): 1693-1709.
[26]	WANG Z Y, LI C, XU H Y, et al. Mamba YOLO: A simple baseline for object detection with state space model[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Washington, DC, USA: AAAI Press, 2025, 39(8): 8205-8213.
[27]	WANG Z Y, ZHENG J Q, ZHANG Y C, et al. Mamba-UNet: UNet-like pure visual mamba for medical image segmentation[PP/OL]. arXiv: 2402.05079, 2024.
[28]	LU W, YANG X, CHEN S B. LWGANet: addressing spatial and channel redundancy in remote sensing visual tasks with light-weight grouped attention[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Washington, DC, USA: AAAI Press, 2026, 40(9): 19845-19853.
[29]	YUAN S, QIN H L, YAN X, et al. ASCNet: asymmetric sampling correction network for infrared image destriping[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 1-15.
[30]	HUANG P M, CHAO I T, HUANG P C, et al. Bevanet: Bilateral efficient visual attention network for real-time semantic segmentation[C]// 2025 IEEE International Conference on Image Processing (ICIP). Piscataway, New Jersey, USA: IEEE, 2025: 2778-2783.
[31]	HUANG X H, LIU S Q, ZHANG K, et al. Reverse convolution and its applications to image restoration[C]// 2025 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway, New Jersey, USA: IEEE, 2025: 10507-10516.
[32]	YAN H L, LI A, ZHANG X T, et al. MobileIE: an extremely lightweight and effective ConvNet for real-time image enhancement on mobile devices[C]// 2025 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway, New Jersey, USA: IEEE, 2025: 21949-21960.
[33]	YANG G Y, WANG Y, SHI D M, et al. Golden cudgel network for real-time semantic segmentation[C]// 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2025: 25367-25376.
[34]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[35]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: Single shot MultiBox detector[C]// Computer Vision-ECCV 2016. Cham, Germany: Springer, 2016: 21-37.
[36]	WANG C Y, YEH I H, MARK LIAO H Y. YOLOv9: Learning what you want toLearn using programmable gradient information[C]// Computer Vision-ECCV 2024. Cham, Switzerland: Springer, 2025: 1-21.
[37]	WANG A, CHEN H, LIU L H, et al. YOLOv10: Real-time end-to-end object detection[C]// Advances in Neural Information Processing Systems 37. Vancouver, BC, Canada. Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2024: 107984-108011.
[38]	KHANAM R, HUSSAIN M. YOLOv11: An overview of the key architectural enhancements[PP/OL]. arXiv: 2410.17725, 2024.
[39]	TIAN Y, YE Q, DOERMANN D. YOLOv12: Attention-centric real-time object detectors[PP/OL]. arXiv: 2502.12524, 2025.
[40]	HUANG S H, LU Z C, CUN X D, et al. DEIM: DETR with improved matching for fast convergence[C]// 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2025: 15162-15171.
[41]	HUANG S H, HOU Y J, LIU L F, et al. Real-time object detection meets DINOv3[PP/OL]. arXiv: 2509.20787, 2025.

Category	Total images	Training/images	Validation/images	Test/images
Alternaria leaf spot	812	568	82	162
Brown spot	790	553	79	158
Frogeye leaf spot	805	564	80	161
Healthy leaf	795	556	80	159
Mosaic	788	552	78	158
Powdery mildew	815	570	82	163
Rust	800	560	80	160
Scab	804	562	81	161

Backbone	P/%	R/%	mAP@0.50/%	mAP@0.75/%	mAP@0.50∶0.95/%	F ₁/%	FLOPs/G	Parameters/M	Model file size/M	FPS
RT-DETR-r18	87.6	78.8	80.7	49.2	49.5	82.9	57.0	19.8	77.0	112.4
MambaYolo	87.2	78.8	82.6	43.6	45.3	82.5	17.2	9.1	34.9	185.2
Mamba	89.6	82.2	84.7	48.2	49.3	85.6	54.5	19.2	74.5	108.7
LWGA	89.4	81.5	84.2	47.4	48.5	85.1	44.7	13.3	52.1	129.9
CNCM	88.2	82.2	84.8	48.0	48.3	85.0	64.5	19.7	76.6	99.0
EVA	88.3	82.1	83.9	47.5	48.8	84.9	48.8	14.5	56.7	123.5
FDEM	89.8	82.4	84.5	50.1	51.0	85.9	47.5	16.4	73.3	120.5

AIFI	P/%	R/%	mAP@0.50/%	mAP@0.75/%	mAP@0.50∶0.95/%	F ₁/%	FLOPs/G	Parameters/M	Model file size/M	FPS
SWSA	84.8	77.5	79.0	46.4	46.9	81.0	55.6	18.3	71.2	115.6
DWMMSA	85.4	78.3	79.8	47.2	47.7	81.7	56.1	18.9	73.5	114.2
BinaryAttn	86.0	79.1	80.4	48.1	48.5	82.4	56.7	19.2	75.1	113.1
WCA	86.9	80.1	81.1	49.1	49.3	83.4	57.0	19.6	76.4	112.4
LRGA	87.9	80.8	81.5	49.4	49.8	84.2	57.2	19.9	77.3	111.8

RepC3	P/%	R/%	mAP@0.50/%	mAP@0.75/%	mAP@0.50∶0.95/%	F ₁/%	FLOPs/G	Parameters/M	Model file size/M	FPS
Converse2DC3^［31］	88.1	79.8	82.7	44.6	46.3	83.7	48.1	18.1	69.5	127.6
MBRConv5C3^［32］	87.8	80.4	82.3	45.8	47.1	83.9	52.4	19.1	73.3	119.4
GCConvC3^［33］	88.2	81.7	82.9	46.7	48.5	84.8	57.0	19.9	76.4	112.3
PDRL	88.1	81.1	81.8	49.6	49.9	84.4	56.7	19.8	76.0	113.0

Model	P/%	R/%	mAP@0.50/%	mAP@0.75/%	mAP@0.50∶0.95/%	F ₁/%	FLOPs/G	Parameters/M	Model file size/M	FPS
RT-DETR-r18	87.6	78.8	80.7	49.2	49.5	82.9	57.0	19.8	77.0	112.4
Faster-RCNN^［34］	81.3	73.7	75.5	44.0	43.5	77.3	370.2	137.1	108.0	28.6
SSD^［35］	87.1	60.0	74.0	44.3	43.9	71.1	62.7	26.3	94.1	97.8
YOLOv5s	86.3	77.6	82.0	49.6	47.8	81.7	16.0	7.2	29.0	219.4
YOLOv8s	86.5	79.0	81.9	49.1	49.4	82.6	28.6	11.2	44.8	176.8
YOLOv9t^［36］	86.9	78.4	82.6	50.3	50.1	82.4	7.7	2.0	8.0	275.6
YOLOv10n^［37］	85.6	76.6	81.4	49.5	50.0	80.8	6.7	2.3	9.2	292.4
YOLOv11n^［38］	86.8	77.5	82.7	50.0	50.3	81.9	6.3	2.6	10.4	301.2
YOLOv12n^［39］	86.1	75.9	80.8	48.0	49.3	80.7	6.5	2.6	10.4	296.7
RT-DETR-r34	89.0	82.1	84.9	51.4	51.3	85.4	95.2	31.8	123.6	73.8
DETRv2-r18	88.7	81.7	84.7	50.9	50.8	85.1	58.2	20.1	78.1	109.8
D-FINE-N	87.2	78.6	83.0	50.5	50.3	82.7	8.4	4.2	16.4	252.7
D-FINE-S	88.5	80.7	84.5	51.5	51.3	84.4	24.6	10.7	42.2	161.3
DEIM^［40］	88.7	78.7	81.9	49.7	49.5	83.4	25.0	10.0	41.2	158.4
DEIMv2^［41］	88.3	79.8	83.1	50.2	49.8	83.8	25.6	9.7	39.8	162.5
FLPNet	89.6	82.7	85.7	52.1	52.0	86.0	47.9	16.5	74.0	119.6