An enhanced vision transformer network for efficient and accurate crop disease detection

Expert Systems with Applications, Год журнала: 2025, Номер unknown, С. 127743 - 127743

Опубликована: Апрель 1, 2025

Язык: Английский

---

A dual-branch model combining convolution and vision transformer for crop disease classification

PLoS ONE, Год журнала: 2025, Номер 20(4), С. e0321753 - e0321753

Опубликована: Апрель 24, 2025

Computer vision holds tremendous potential in crop disease classification, but the complex texture and shape characteristics of diseases make classification challenging. To address these issues, this paper proposes a dual-branch model for which combines Convolutional Neural Network (CNN) with Vision Transformer (ViT). Here, convolutional branch is utilized to capture local features while handle global features. A learnable parameter used achieve linear weighted fusion two types An Aggregated Local Perceptive Feed Forward Layer (ALP-FFN) introduced enhance model’s representation capability by introducing locality into encoder. Furthermore, constructs lightweight block using ALP-FFN self-attention mechanism reduce parameters computational cost. The proposed achieves an exceptional accuracy 99.71% on PlantVillage dataset only 4.9M 0.62G FLOPs, surpassing state-of-the-art TNT-S (accuracy: 99.11%, parameters: 23.31M, FLOPs: 4.85G) 0.6%. On Potato Leaf dataset, attains 98.78% accuracy, outperforming advanced ResNet-18 98.05%, 11.18M, 1.82G) 0.73%. effectively advantages CNN ViT maintaining design, providing effective method precise identification diseases.

Язык: Английский

---

Attention-based unsupervised prompt learning for SAM in leaf disease segmentation

Knowledge-Based Systems, Год журнала: 2025, Номер unknown, С. 113652 - 113652

Опубликована: Май 1, 2025

Язык: Английский

---

LKCAFormer: A Lightweight Transformer with Large-Kernel Cooperative Attention for the Segmentation of Field Maize Leaf Diseases

Опубликована: Май 15, 2025

In smart agriculture, segmentation models are essential for the early and accurate detection of diseases. However, complex backgrounds diverse diseases on maize leaves present significant challenges. Although current have improved, these advancements often lead to larger model sizes higher computational demands, making them difficult deploy hardware with limited resources. To overcome issues, we propose a new lightweight network called LKCAFormer. This is specifically designed leaf disease built upon coordinated attention mechanism cross-scale large-kernel convolutions. Our approach introduces Large-Kernel Convolution Cooperative Attention (LK-COA) module, which uses convolutions extract global features cooperative capture fine details small spots. combination enhances spots reduces errors caused by spot adhesion. Additionally, CSDecoder effectively fuses shallow features, rich in edge detail information, deeper semantic produce precise results. Experimental results three datasets demonstrate that our method outperforms existing techniques, confirming its effectiveness pathological analysis

Язык: Английский

---

PID-Former: A Triple Stream Network for Real-Time Coal Flow Segmentation

Research Square (Research Square), Год журнала: 2025, Номер unknown

Опубликована: Май 19, 2025

Coal flow segmentation is the basis for detecting coal pile and deviation on conveyor belts, real-time performance accuracy of methods are key factors affecting above functions. However, segmenting from belts with similar color characteristics complex lighting a challenging task. To address issues, we proposed triple stream semantic model PID-Former, which consists P, I D branches, achieves segmentation. Firstly, I-Branch designed extracting context information based lightweight transformer framework, P-Branch D-Branch constructed by use Inverted Residual Module (IRM) detail boundary respectively. Secondly, Context Information Fill (CIFM) Edge Fusion (EFM) to combine different levels’ supplement information. At same time, strengthens in features through Spatial Enhancement Block (SEB). Finally, PID fuse detail, extracted I, branches. Experimental results show that, PID-Former achieved 96.99% mIoU 98.68% mPA dataset, reached speed 204.1FPS 67.1FPS RTX3090 GPU embedded platform NVIDIA Jetson TX2 Compared State-of-Arts, better trade-off between inference speed.

Язык: Английский

---

PMJDM: a multi-task joint detection model for plant disease identification

Frontiers in Plant Science, Год журнала: 2025, Номер 16

Опубликована: Май 22, 2025

Introduction Plant disease detection is critical for ensuring agricultural productivity, yet traditional methods often suffer from inefficiencies and inaccuracies due to manual processes limited adaptability. Methods This paper presents the PlantDisease Multi-task Joint Detection Model (PMJDM), which integrates an enhanced ConvNeXt-based shared feature extraction, a texture-augmented N-RPN module with HOG/LBP metrics, multi-task branches simultaneous plant species classification detection, CRF-based post-processing spatial consistency. A dynamic weight adjustment mechanism also employed optimize task balance improve robustness. Results Evaluated on 26,073-image dataset, PMJDM achieves 71.84% precision, 61.96% recall, 61.83% mAP50, surpassing Faster - RCNN (51.49% mAP50) YOLOv10x (59.52% by 10.34% 2.31%, respectively. Discussion The superior performance of driven synergy texture region proposals, offering efficient solution precision agriculture.

Язык: Английский

---

LT-DeepLab: an improved DeepLabV3+ cross-scale segmentation algorithm for Zanthoxylum bungeanum Maxim leaf-trunk diseases in real-world environments

Frontiers in Plant Science, Год журнала: 2024, Номер 15

Опубликована: Окт. 22, 2024

Introduction Zanthoxylum bungeanum Maxim is an economically significant crop in Asia, but large-scale cultivation often threatened by frequent diseases, leading to yield declines. Deep learning-based methods for disease recognition have emerged as a vital research area agriculture. Methods This paper presents novel model, LT-DeepLab, the semantic segmentation of leaf spot (folium macula), rust, frost damage (gelu damnum), and diseased leaves trunks complex field environments. The proposed model enhances DeepLabV3+ with innovative Fission Depth Separable CRCC Atrous Spatial Pyramid Pooling module, which reduces structural parameters module improves cross-scale extraction capability. Incorporating Criss-Cross Attention Convolutional Block Module provides complementary boost channel feature extraction. Additionally, deformable convolution low-dimensional features, Fully Network auxiliary header integrated optimize network enhance accuracy without increasing parameter count. Results LT-DeepLab mean Intersection over Union (mIoU) 3.59%, Pixel Accuracy (mPA) 2.16%, Overall (OA) 0.94% compared baseline DeepLabV3+. It also computational demands 11.11% decreases count 16.82%. Discussion These results indicate that demonstrates excellent capabilities environments while maintaining high efficiency, offering promising solution improving management efficiency.

Язык: Английский

---

EFS-Former: An Efficient Network for Fruit Tree Leaf Disease Segmentation and Severity Assessment

Agronomy, Год журнала: 2024, Номер 14(9), С. 1992 - 1992

Опубликована: Сен. 2, 2024

Fruit is a major source of vitamins, minerals, and dietary fiber in people’s daily lives. Leaf diseases caused by climate change other factors have significantly reduced fruit production. Deep learning methods for segmenting leaf can effectively mitigate this issue. However, challenges such as folding, jaggedness, light shading make edge feature extraction difficult, affecting segmentation accuracy. To address these problems, paper proposes method based on EFS-Former. The expanded local detail (ELD) module extends the model’s receptive field expanding convolution, better handling fine spots reducing information loss. H-attention reduces computational redundancy superimposing multi-layer convolutions, improving filtering. parallel fusion architecture utilizes different intervals convolutional neural network (CNN) Transformer encoders, achieving comprehensive fusing detailed semantic channel spatial dimensions within (FFM). Experiments show that, compared to DeepLabV3+, achieves 10.78%, 9.51%, 0.72%, 8.00% higher scores mean intersection over union (mIoU), pixel accuracy (mPA), (Acc), F_score, respectively, while having 1.78 M fewer total parameters 0.32 G lower floating point operations per second (FLOPS). Additionally, it calculates ratio area occupied spots. This also effective calculating disease period analyzing diseased method’s overall performance evaluated using mIoU, mPA, Acc, F_score metrics, 88.60%, 93.49%, 98.60%, 95.90%, respectively. In summary, study offers an efficient accurate tree spot segmentation, providing solid foundation precise analysis leaves spots, supporting smart agriculture precision pesticide spraying.

Язык: Английский

---

Reformer: Re-parameterized kernel lightweight transformer for grape disease segmentation

Expert Systems with Applications, Год журнала: 2024, Номер 265, С. 125757 - 125757

Опубликована: Дек. 9, 2024

Язык: Английский

---