Accurate, real-time identification of tea leaf diseases is important for preserving crop quality and reducing economic losses. Traditional image-processing and shallow learning approaches (color histograms, texture descriptors, SVMs) have been largely superseded by deep convolutional neural networks because CNNs extract hierarchical features that suit complex field environments. However, deploying high-performing networks on edge hardware faces a persistent trade-off between detection accuracy and computational cost. This study proposes CA-YOLOv8n, a lightweight detection model tailored to tea disease diagnosis that aims to improve spatial sensitivity to subtle lesions while reducing model size and inference latency for on-site use.
The work addresses three interconnected problems: (1) spatial information dilution in deep architectures that leads to missed small or early lesions, (2) the limitations of channel-only attention mechanisms that discard positional cues, and (3) the need for compact, interpretable models with Params < 5M that can run on commodity CPUs in agricultural settings. The authors position their contributions within the established progression from early manual feature engineering systems to modern one-stage detectors such as the YOLO family.
Coordinate Attention (CA) is adopted to overcome the loss of position-sensitive cues that occurs in many channel-based attention blocks (for example, Squeeze-and-Excitation). Unlike purely channel-wise recalibration, CA encodes directional and positional information into channel attention, enabling the network to better preserve spatial dependencies. In the context of tea leaf disease detection, this is intended to help the model localize small, irregularly shaped lesions—such as early-stage white spots—that are easily suppressed as noise in lightweight backbones.
The integration of Coordinate Attention into the YOLOv8n backbone is aimed at enhancing the model's ability to discriminate subtle pathological features from complex backgrounds like canopy occlusion and variable illumination commonly found in field-collected images.
To reduce computational overhead without sacrificing multi-scale detection capability, the authors introduce a Path-Decoupling strategy that streamlines feature fusion within the network. The Path-Decoupling approach restructures how features from different scales are combined, simplifying the network topology and lowering resource requirements (parameters and FLOPs). This architectural change targets the key practical constraint for edge deployment: minimizing model complexity while maintaining detection sensitivity across lesion sizes.
The Path-Decoupling strategy is described as complementary to the CA module: Path-Decoupling reduces redundant computations in feature fusion, while CA preserves important spatial cues within the learned representations.
The study evaluates the proposed model against a baseline YOLOv8n implementation. Evaluation metrics reported include mean Average Precision at 50% overlap (mAP@50), parameter count, floating-point operations per second (FLOPs), Recall, and CPU inference latency. The dataset used for training and testing is publicly available on figshare via the stated DOI. Additional implementation and deployment details were provided in the full paper; specific training hyperparameters and hardware configurations were reported in the source article but are not restated here.
Reported performance improvements for CA-YOLOv8n relative to the baseline YOLOv8n include:
These metrics indicate that the proposed model achieves high detection precision while substantially lowering model complexity.
Beyond model design and evaluation, the authors integrated CA-YOLOv8n into a diagnostic platform with an automated reporting interface. The platform demonstration emphasized practical feasibility: the model can perform real-time tea disease identification on commodity CPU hardware typical of resource-constrained agricultural environments.
On the platform, the proposed model achieved a reported 23.2% reduction in CPU inference latency compared to baseline YOLOv8n, supporting claims that the architectural modifications enable faster on-site processing and more responsive diagnostic reporting.
Strengths:
Limitations and considerations noted by the study:
Implications:
The authors present CA-YOLOv8n, a lightweight, spatially aware detection model that integrates Coordinate Attention with a Path-Decoupling strategy to improve detection of multi-scale tea diseases while lowering computation and latency. The reported mAP@50 is 98.89% with parameter and FLOPs reductions of 32.6% and 24.1%, respectively; Recall improved by 1.21 percentage points and CPU inference latency dropped by 23.2% on the diagnostic platform compared to YOLOv8n.
All data underlying the study are publicly available on figshare at the DOI provided in the source article. The authors reported no specific funding and declared no competing interests.