Journal of Electronic Science and Technology, Volume. 23, Issue 1, 100300(2025)
Grape Guard: A YOLO-based mobile application for detecting grape leaf diseases
Figures & Tables(16)
Fig. 2. Grape leaf images of each label (from the grape leaf disease dataset).
Fig. 4. Precision confidence curve of YOLOv8 with image size 640×640.
Fig. 5. Recall confidence curve of YOLOv8 with image size 640×640.
Fig. 6. Precision confidence curve of YOLOv5 with image size 640×640.
Fig. 7. Recall confidence curve of YOLOv5 with image size 640×640.
Table 1. Research matrix
View tableView in ArticleTable 1. Research matrix
Author Dataset Model Results Contribution Lu et al. [ 3 ]GLDP12k Ghost convolution enlightened the transformer Accuracy: 98.14% It proposed a transformer-based network that generates intermediate feature maps. Li et al. [ 4 ]Two small-scale datasets DCT Accuracy: 89.19% Accuracy: 93.92% Proposed a DCT model where the transformer serves as the backbone of this model. Lu et al. [ 5 ]Washington State University (WSU) Roza Experimental Orchards, Prosser, WA Swin-T-YOLOv5 mAP: 97% Developed Swin-T-YOLOv5 by architecturally integrating YOLOv5 and Swin-transformer detectors. Praveen et al. [ 6 ]Plant Village Dataset YOLO-X with with SE, ECA, and CBAM attention Precision: 89.77%, Recall: 86.97%, F1 Score: 85.91%, and mAP: 88.96% Combined attention techniques such as CBAM, SE, and ECA with YOLO to improve GLDD. Xie et al. [ 7 ]GLDD dataset Faster DR-IACNN mAP: 81.1% Enhanced the Faster R-CNN model with Inception-v1, Inception-ResNet-v2 modules, and SE blocks to improve feature extraction. Kong et al. [ 8 ]Jilin Agricultural University YOLOv5 network and transformer module mAP@0.5: 84.3% YOLOv5 network and transformer module were used to detect targets. Shaheed et al. [ 9 ]PlantVillage repository Efficient RMT-Net Accuracy: 97.65% Accuracy: 99.12% A combinational model with vision transformer and ResNet-50 where efficient RMT-Net, distinct features are extracted using the CNN model. Xia et al. [ 10 ]Research Institute of Pomology of Chinese Academy of Agricultural Sciences in Xingcheng and the Beijing Vocational College of Agriculture in Beijing, China MTYOLOX AP50: 83.4% AR50: 93.3% Developed ST-PAFPN module and DAT-Darknet module, which are embedded into the backbone and neck of the network based on multiple self-attention mechanisms. Leng et al. [ 11 ]NLB dataset YOLOv5 mAP@0.5: 87.5% Introduced the feature restructuring and fusion module, which focuses on retaining critical information during downsampling. Lu et al. [ 12 ]WGISD CMA-YOLO Precision: 89.6% F1 score: 86.5% AP: 90.2% Introduced a YOLOv5-based model that integrates dual-stream data loading, mosaic augmentation, global self-attention, and a CMA-C3 module to enhance grapefruit detection accuracy. Feng et al. [ 13 ]Xiaotangshan National Precision Agriculture Demonstration Base YOLOv5s+BiCMT Accuracy: 99.23% Precision: 97.37% Sensitivity: 97.54% Specificity: 99.54% Proposed YOLOv5 for region detection and a BiCMT classifier for feature fusion. Li et al. [ 14 ]Dangshan County, Suzhou City, Anhui Province, China YOLOv5s-FP AP: 96.12% Developed YOLOv5s-FP, which utilises a modified CSP module with a transformer encoder for global feature extraction and attentional feature fusion. Jiang et al. [ 15 ]/ Efficient LC3Net model AP: 92.29% Proposed the Retinex algorithm for contrast enhancement and LC3Net model, with image normalization and reduced down-sampling frequency. Sun et al. [ 16 ]PlantVillage dataset SE-VIT hybrid network Accuracy: 97.26% Developed SE-VIT hybrid network where the SE attention module enhances inter-channel weight learning in ResNet-18. Huang et al. [ 17 ]/ YOLO-EP algorithm, based on YOLOv5 AP@0.5: 88.6% Precision: 85.1% Recall: 82.6%. Introduced the YOLO-EP algorithm, utilizing transposed convolution and attention algorithms. Thai et al. [ 18 ]Cassava Leaf Disease Dataset Least important attention pruning (LeIAP) algorithm / Developed LeIAP algorithm to select each layer’s most critical attention heads in the transformer model. Chen et al. [ 19 ]/ ESP-YOLO mAP: 98.3% Integrated YOLO with advanced techniques like ELSAN, SE, and PConv to improve the accuracy and efficiency of table grape detection. Liu et al. [ 20 ]RGB Grape Data -North China FRT-YOLO mAP: 90.67% Developed FTR-YOLO, a real-time and lightweight model for detecting grape diseases.
Table 2. Description of the grape leaf diseases
View tableView in ArticleTable 2. Description of the grape leaf diseases
Labels Characteristics Black measle Black measle is a fungal disease that causes small, reddish spots on leaves that eventually turn brown or black [ 23 ]. If left untreated, black measles can severely reduce grape yields and quality.Black rot The fungal pathogen Guignardia Bidwell [ 24 ] causes the fungus Black Rot. Black Rot typically appears on grape leaves as small, yellow spots that gradually enlarge and turn brown or black [25 ]. Disease-affected areas may become necrotic, causing the leaves to wither and die.Blight fungus It refers to various fungal diseases that cause blighting symptoms on leaves. It is characterised by irregular lesions, spots, or imperfections on the leaves ranging from brown to black [ 26 ].Healthy It is the standard, unaffected state of grape leaves. The leaves of healthy grapes are typically green and free of spots, lesions, or discolourations.
Table 3. Parameter configuration for the experiment.
View tableView in ArticleTable 3. Parameter configuration for the experiment.
Image size Model wights Batch size Epoch Model 640 YOLOv8l.pt 16 100 YOLOv8 320 YOLOv8l.pt 16 100 256 YOLOv8l.pt 16 100 128 YOLOv8l.pt 16 100 640 YOLOv5l.pt 16 100 YOLOv5 320 YOLOv5l.pt 16 100 256 YOLOv5l.pt 16 100 128 YOLOv5l.pt 16 100
Table 4. Classification report of YOLOv8 and YOLOv5 with different image sizes.
View tableView in ArticleTable 4. Classification report of YOLOv8 and YOLOv5 with different image sizes.
YOLOv8 YOLOv5 Classes Precision Recall mAP50 mAP50-95 Training time (H) Precision Recall mAP50 mAP50-95 Instances Training time (H) Image Size 640 All 0.999 1.000 0.995 0.88 0.883 0.996 0.995 0.995 0.86 319 0.545 Black measles 0.999 1.000 0.995 0.871 1.000 0.997 0.995 0.857 80 Black fot 1.000 1.000 0.995 0.866 0.998 1.000 0.995 0.844 89 Blight fungus 0.999 1.000 0.995 0.887 1.000 0.985 0.995 0.875 77 Healthy 0.999 1.000 0.995 0.897 0.987 1.000 0.995 0.882 73 Image Size 320 All 0.966 0.982 0.993 0.858 0.861 0.998 1 0.995 0.869 319 0.516 Black measles 0.967 0.963 0.994 0.842 0.997 1 0.995 0.869 80 Black rot 0.956 0.979 0.992 0.846 1 1 0.995 0.86 89 Blight fungus 0.964 0.987 0.994 0.877 0.997 1 0.995 0.872 77 Healthy 0.977 1 0.994 0.868 0.998 1 0.995 0.873 73 Image size 256 All 0.994 0.997 0.995 0.88 0.833 0.996 0.994 0.995 0.863 319 0.432 Black measles 0.99 1.000 0.995 0.88 0.999 1.000 0.995 0.856 80 Black rot 1.000 0.992 0.995 0.866 1.000 1.000 0.995 0.853 89 Blight fungus 1.000 0.995 0.995 0.889 1.000 0.976 0.995 0.868 77 Healthy 0.987 1.00 0.995 0.885 0.987 1.000 0.995 0.875 73 Image size 128 All 0.998 1.000 0.995 0.873 0.792 0.996 0.998 0.995 0.835 319 0.345 Black measles 0.999 1.000 0.995 0.861 1.000 0.990 0.995 0.832 80 Black rot 1.000 0.999 0.995 0.862 0.989 1.000 0.995 0.845 89 Blight fungus 0.998 1.000 0.995 0.88 0.999 1.000 0.995 0.841 77 Healthy 0.997 1.000 0.995 0.887 0.995 1.000 0.995 0.821 73
Table 5. Input-Output shape and data type of the TFLite model.
View tableView in ArticleTable 5. Input-Output shape and data type of the TFLite model.
Input Output Type [1, 640×640, 3] [1, 25200 , 9]Shape NumPy.float32 NumPy.float32 Data Type
Tools
Get Citation
Copy Citation Text
Sajib Bin Mamun, Israt Jahan Payel, Md. Taimur Ahad, Anthony S. Atkins, Bo Song, Yan Li. Grape Guard: A YOLO-based mobile application for detecting grape leaf diseases[J]. Journal of Electronic Science and Technology, 2025, 23(1): 100300
Paper Information
Category:
Received: Apr. 1, 2024
Accepted: Jan. 5, 2025
Published Online: Apr. 7, 2025
The Author Email: Md. Taimur Ahad (MdTaimur.Ahad@unisq.edu.au)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20