Multiscale Tea Disease Detection with Channel–Spatial Attention

Abstract

Tea disease detection is crucial for improving the agricultural circular economy. Deep learning-based methods have been widely applied to this task, and the main idea of these methods is to extract multiscale coarse features of diseases using the backbone network and fuse these features through the neck for accurate disease detection. This paper proposes a novel tea disease detection method that enhances feature expression of the backbone network and the feature fusion capability of the neck: (1) constructing an inverted residual self-attention module as a backbone plugin to capture the long-distance dependencies of disease spots on the leaves; and (2) developing a channel–spatial attention module with residual connection in the neck network to enhance the contextual semantic information of fused features in disease images and eliminate complex background noise. For the second step, the proposed channel–spatial attention module uses Residual Channel Attention (RCA) to enhance inter-channel interactions, facilitating discrimination between disease spots and normal leaf regions, and employs spatial attention (SA) to enhance essential areas of tea diseases. Experimental results demonstrate that the proposed method achieved accuracy and mAP scores of 92.9% and 94.6%, respectively. In particular, this method demonstrated improvements of 6.4% in accuracy and 6.2% in mAP compared to the SSD model.

1. Introduction

Tea is considered a very valuable green cash crop due to its profound health benefits for humans [1], making the tea industry an important part of the agricultural circular economy. Tea disease (such as leaf blight, rust algae disease, and red spot disease) annually lead to a decrease of approximately 20% in tea production, leading to declines in yield and quality as well as resource wastage [2]. Therefore, how to develop a precise and timely tea disease detector is meaningful for early warning and accurate control of diseases, and it can enhance control efficiency and reduce pesticide use, aligning with the principles of reducing chemical inputs and protecting the environment in the agricultural circular economy.

Traditional tea disease detection primarily relies on artificial methods [3,4]. However, these methods exhibit a high misdiagnosis rate and low detection efficiency. Many convolution-based deep learning methods for tea disease detection have been proposed, for instance, the YOLO series [5,6,7] and other object detection methods [8,9]. However, these object detection methods of tea disease still have limitations in integrating global information and handling long-range dependencies.

The attention mechanism can effectively overcome the limitations of convolution-based deep learning methods [10], and existing attention-based techniques can be roughly grouped into three groups: channel attention-based, spatial attention-based, and self-attention-based methods [11,12]. Channel-based techniques enhance the model’s focus on important features by dynamically adjusting the weight of each channel [13]. Liu et al. [14] introduced a channel attention mechanism that assigns weights to individual channels and multiplies them with the original disease image to emphasize significant features while suppressing irrelevant background noise. Zhang et al. [15] presented a multi-channel automatic directional circulation mechanism to extract and reuse multiscale features to overcome challenges associated with the similarity among tomato leaf diseases. Gao et al. [16] introduced a highly efficient crop disease recognition model to improve the accuracy of plant disease recognition while minimizing the extraction of extraneous features. Wenxia Bao et al. [17] proposed an improved target detection and recognition network, AX-RetinaNet, for the automatic detection and recognition of tea leaf diseases in natural scene images. The attention module assigns adaptively optimized weights to each feature map channel, enabling the network to select more effective features and reduce the interference of redundant features. Although these methods effectively tackle redundancy and background noise at the channel level, they encounter challenges in extracting spatial position information and capturing inter-regional relationships.

Spatial attention-based methods aim to analyze and highlight particular regions and locations in an image to enhance the understanding of local structure and contextual information [18]. Wang et al. [19] utilized spatial attention to establish correlations between different regions, thereby enhancing the importance of critical regions in the feature map. Tang et al. [20] incorporated a multiscale spatial attention block into fully convolutional neural networks to address the issue of spatial information loss. Wang et al. [21] proposed an attention model that incorporates correlation spatial attention, increasing the importance of critical regions in the feature map by establishing correlations between different regions. Xie et al. [22] proposed a tea disease detection model based on YOLOv8s, integrating deformable convolutions, attention mechanisms, and an improved spatial pyramid pooling module to enhance the model’s capability to handle complex targets and challenging backgrounds by reducing interference from irrelevant factors and enabling effective multi-feature fusion. Although these methods effectively extract spatial information, they do not incorporate spatial attention for capturing multiscale features.

Self-attention can be seen as a special form of spatial attention that captures spatial relationships among individual pixels, allowing for efficient feature representation [23,24]. Therefore, introducing self-attention mechanisms can enhance multiscale feature extraction [25,26]. Stephen et al. [27] incorporated self-attention into the ResNet18 and ResNet34 architectures to enhance the process of feature selection. Zeng et al. [28] introduced a Self-focused Convolutional Neural Network (SACNN) that incorporates a self-attention mechanism to extract relevant features from crop disease spots for disease identification purposes. The above studies demonstrate that the self-attention mechanism is very effective in enhancing feature extraction in convolutional neural networks.

Based on the above analysis, integrating self-attention into convolutional neural networks for multiscale feature extraction is feasible. Furthermore, spatial and channel attention effectively reduce noise and redundancy. Therefore, this paper proposes a tea disease detection method based on channel–spatial attention (TCSA). The primary contributions are summarized as follows:

• TCSA enhances the backbone network’s ability to capture long-range dependencies of disease spots on leaves using inverted residual self-attention modules. Additionally, TCSA improves feature fusion in the neck with channel–spatial attention modules, enriching contextual semantic information of diseased regions and eliminating complex background noise.
• This paper designs a channel attention module, named Residual Channel Attention (RCA). RCA uses residual connections to enhance channel interactions and employs two pooling techniques to capture global information from each channel, thereby improving the distinction between diseased and healthy tea leaves.
• This paper proposes a spatial attention module called spatial attention (SA). SA optimizes attention computation to reduce computational complexity and introduces Depth-Wise Convolution (DW-Conv) to focus on more informative disease regions, further reducing the computational load.

We organize ablation experiments and comparative analyses to demonstrate the effectiveness of the proposed method. Experimental results show that the detection accuracy of the proposed method is improved in multiscale and complex background scenes.

This paper is structured as follows: Section 2 introduces the dataset utilized in this paper. Section 3 introduces the structure of TCSA and the channel–spatial attention module. Section 4 describes the configuration of the experimental equipment, some experimental parameters, the results of the ablation and comparative experiments, and the performance of TCSA on different diseases. Section 5 concludes with recommendations and future directions.

2. Materials

2.1. Disease Dataset

The dataset of tea used in the deep learning data modeling was acquired from Anhui Agricultural University [29]. This dataset consists of 776 tea disease images with a size of 906 × 600 pixels, and each image can be grouped into one of the six common tea disease categories, i.e., tea algae leaf spot (Als), tea cake (Tc), tea cloud leaf blight (Clb), tea exobasidium blight (Eb), tea red rust (Tr), and tea red scab (Rs), as shown in Figure 1. Due to the original dataset without the bounding boxes of the tea disease targets, we used a tool called labelimg for labeling the images to assess the model’s performance on the problem of tea leaf pest and disease detection.

2.2. Data Augmentation

To enhance the model’s robustness and generalization capability, data augmentation techniques including image flipping, horizontal and vertical mirroring, and adding noise have been employed in this method, as shown in Figure 2. Additionally, the number of 776 tea disease images increased to 3676 after data augmentation, and the enhanced dataset was divided into a training set, validation set, and test dataset in the ratio of 8:1:1.

3. Methodology

To solve the tea disease detection problems of varying disease scales, occlusions of tea leaves, and the complexity of backgrounds in natural scenes, we propose a tea disease detection with channel–spatial attention (TCSA), consisting of three parts: backbone, neck, and head, as shown in Figure 3.

The backbone composed of the multiscale convolutional block and self-attention mechanism captures features with multiple scales. Then, the neck uses a top-down/bottom-up fashion to fuse multiscale features generated from the backbone to aid in capturing both the global information and detailed features of tea leaf diseases, such as distribution patterns, small spots, and edge details. In addition, we propose a novel channel attention module and a spatial attention module, and incorporate them as plugins after each feature fusion operation to eliminate the background noise and highlight the disease features. Finally, the head consists of three parallel convolution branches to detect tea diseases with multiple sizes.

3.1. Backbone

The backbone is used to extract multiscale coarse features of diseases. There are many commonly used models, such as ResNet-101 [30] and Darknet53 [31], that can be selected as backbones for extracting features. We chose CSPDarkNet53 [32], which has demonstrated excellent results in many classic deep learning-based methods [33,34], as our backbone framework to extract multiscale features of tea diseases, due to CSPDarkNet53’s low computational complexity and strong adaptability. Additionally, we introduce Inverted Residual Mobile self-attention Block (IRMB) [35] as plugins inserted into CSPDarkNet53 to solve the limitation of CSPDarkNet53 in integrating global information and handling long-range dependencies, as shown in Figure 3. From Figure 3, we insert IRMB at four dimensions of CSPDarknet53: 160 × 160 × 128, 80 × 80 × 256, 40 × 40 × 512, and 20 × 20 × 1024. The disease features extracted by IRMB at the dimensions of 80 × 80 × 256, 40 × 40 × 512, and 20 × 20 × 1024 are then output to the neck.

Figure 4 shows the details of the proposed Inverted Residual Mobile self-attention Block (IRMB). IRMB combines self-attention with Depth-Wise Convolution (DW-Conv). The 1 × 1 convolutions compress and expand the number of channels to optimize computational efficiency, and the 3 × 3 DW-Conv captures spatial features, while the attention mechanism captures global dependencies between features. We inserted IRMB into the CSPDarkNet53 structure, which not only optimizes computational efficiency but also enhances the model’s ability to extract multiscale features of tea diseases.

3.2. Neck

The neck repeatedly fuses multiscale tea disease feature maps through a top-down/bottom-up fashion. In addition, we utilize Residual Channel Attention (RCA) and spatial attention (SA) in the neck to enhance the contextual semantics of fused features of the disease image, and to eliminate redundancy and noise from complex backgrounds, as shown in Figure 3.

3.2.1. Residual Channel Attention Module

Channel attention automatically highlights relevant feature channels while suppressing irrelevant ones [36]. Therefore, we propose a novel Residual Channel Attention (RCA) module which uses Global Average Pooling (GAP) and Global Maximum Pooling (GMP) to retain relevant channel information. Specifically, GAP computes the mean value of each feature channel, providing a generalized summary of the feature information, while GMP identifies the maximum value within each feature channel, highlighting the most prominent features. Additionally, residual connections are used to strengthen the interactions between channels, enhancing the overall feature representation.

Figure 5 shows the details of the proposed Residual Channel Attention (RCA). Given the input x, RCA uses Global Average Pooling (GAP) and Global Maximum Pooling (GMP) followed by a shared Multi-Layer Perceptron (MLP) with a ReLU operation and a convolution with a kernel size of 1 × 1 to capture the global information of each channel, denoted by x₁ and x₂. Then, RCA aggregates x₁ and x₂ using element-wise addition followed by the sigmoid to obtain channel attention weights that are used for reweighting the input feature maps. Therefore, RCA is formulized as follows:

$$\theta =Sigmoid\left(ReLU\left(MLP\left(GAP\left(x\right)\right)\right)+ReLU\left(MLP\left(GMP\left(x\right)\right)\right)\right)$$

(1)

$$FC=x\otimes \theta \oplus x$$

(2)

where $\otimes$ means element-wise multiplication, $\oplus$ means element-wise addition, and FC means channel feature maps.

3.2.2. Spatial Attention Module

In this section, we propose a novel spatial attention (SA) module with Self-Attention Block (SAB) as the plugin to focus on disease areas that may be occluded by other leaves and reduce spatial dimensions to optimize computational complexity, thereby addressing the issue of extensive leaf occlusion in complex backgrounds. Note: in the field of visual research, the self-attention mechanism can be considered, to some extent, as a form of spatial attention mechanism. Figure 6a shows the overall structure of the proposed SA. SA includes the Self-Attention Block (SAB), Depth-Wise Convolution (DW-Conv), residual connections, and layer normalization. Figure 6b illustrates the details of how SAB reduces the spatial dimensions. SAB includes a Spatial Decrease module (SD) and Multi-Head Attention (MHA). SAB receives Query, Key, and Value as input. Then, SD conducts convolution operations on the Key and Value of kernel size N × N. The Key and Value space dimension of H × W is reduced to HW/N². The formula is as follows:

$$SD\left(x\right)=Norm\left(Conv\left(Key,Value,N\right)\right)$$

(3)

where Conv(Key, Value, N) is a convolution operation with kernel size N × N on the input Key and Value. Norm(·) refers to layer normalization.

Furthermore, DW-Conv can be regarded as local attention, which makes the model pay attention to the adjacent features in the space, to ensure that the model pays attention to the overall disease area while enhancing the local disease detection ability in tea diseases.

Hence, the output of spatial attention is as follows:

$$FS=DW-Conv\left(F{C}^{\prime }\right)+F{C}^{\prime }$$

(4)

$$F{C}^{\prime }=SAB\left(Norm\left(FC\right)\right)+Norm\left(FC\right)$$

(5)

where FS is the spatial feature map.

Finally, the feature map fused through the residual connection is expressed as follows:

$$FSC=FS\otimes x$$

(6)

where FSC is the feature map after the final fusion.

3.3. Head

The head aims to identify tea diseases at different scales. In this paper, we utilize three parallel convolutions to better capture the features of objects at different scales and improve the detection performance, as shown in Figure 3. Three detection heads with sampling ratios of 1/8, 1/16, and 1/32 are used, corresponding to small object detection, medium object detection, and large object detection, respectively.

3.4. Loss Function

The loss function of the proposed model primarily comprises VariFocal Loss (VFL) [37], CIoU loss, and Distribution Focal Loss (DFL) [38].

The main improvement of the VFL is the proposal on asymmetric weighting operations. The mathematical formulation of this loss function is delineated as follows:

$$VFL\left(p,q\right)=\left\{\begin{array}{cc}-q\left(qlog\left(p\right)+\left(1-q\right)log\left(1-p\right)\right)& q>0\\ -\alpha p^{g}log(1-p)& q=0\end{array}\right.$$

(7)

where p represents the predicted Iou-Aware Classification Score (IACS), and q represents the target score.

DFL optimizes the probabilities of the two locations closest to the target label y using cross-entropy, enabling the network to quickly focus on the distribution near the target location.

$$DFL\left(S_i+S_{i+1}\right)=-\left(\left(y_{i+1}-y\right)log(S_i)+\left(y-y_i\right)log(S_{i+1})\right)$$

(8)

$$S_i={\displaystyle \frac{y_{i+1}-y}{y_{i+1}-y_i}}$$

(9)

$$S_{i+1}={\displaystyle \frac{y-y_i}{y_{i+1}-y_i}}$$

(10)

where y represents the label, and y_i and y_i+1 represent the nearest two to y (y_i ≤ y ≤ y_i+1).

CIoU [39] loss is employed to address the issue of mutual occlusion between leaves and a better return to the bounding box.

$$CIoU=IoU-{\displaystyle \frac{{\rho }^2\left(b,b^{gt}\right)}{s^2}}-\alpha \upsilon$$

(11)

where α is the weight coefficient, b and b^gt represent the center point of the predicted and actual frames, respectively, and ν is used to represent the similarity of the aspect ratio.

4. Experiments

4.1. Dataset and Experimental Setup

To eliminate variations in experimental data due to differing conditions, all experiments in this study were conducted under the same hardware and software environments. We used NVIDIA A100 80 GB PCIe GPUs (NVIDIA, Santa Clara, CA, USA) and the Linux operating system. The programming language was Python 3.9, and the deep learning framework was PyTorch 1.8.2. The training was conducted from scratch without using any pre-trained weights. During training, the image size was resized to 640 × 640, the optimizer chosen was Adam, the batch size was set to 32, the initial learning rate was 0.001, and the number of epochs was set to 300.

4.2. Evaluation Metrics

In this study, mean average precision (mAP) is employed for evaluating the proposed methods, which are as follows:

$$mAP={\displaystyle \frac{{\displaystyle \sum _1^{k}AP\left(i\right)}}{k}}$$

(12)

where k is the number of classes in the test set, and AP(i) is the average precision of the model for the i-th class of targets.

The computation of mAP encompasses average precision (AP). In our evaluation, the AP is formulated as follows:

$$AP={\displaystyle \underset{0}{\overset{1}{\int }}P\left(r\right)}dr$$

(13)

P is the percentage of actual target instances correctly predicted by the model among all the predicted target instances. The calculation of precision is as follows:

$$P={\displaystyle \frac{TP}{TP+FP}}$$

(14)

R denotes the percentage of recognized targets by the model among the actual target instances. R is computed using the following formula:

$$R={\displaystyle \frac{TP}{TP+FN}}$$

(15)

where TP refers to correctly identifying the target instance, and FP represents an instance where the target was incorrectly identified. TN means accurate identification of the background instance. FN indicates instances where the background was misidentified and is highlighted when the object was misclassified as background.

4.3. Ablation Study

Ablation experiments were conducted to evaluate the efficacy and contribution of the proposed blocks to overall performance. Each attention mechanism block was incrementally incorporated into the architecture based on the CSPDarkNet53 backbone.

Table 1. Results of ablation experiments with different models.

Backbone (CSPDarkNet53)	IRMB	RCA	SA	mAP
√				93.3
√	√			93.5
√		√		93.7
√			√	94.1
√	√	√		93.9
√	√		√	94.4
√	√	√	√	94.6

shows that, overall, the performance of TCSA improves with the addition of more sub-modules and TCSA with all sub-modules achieves the highest accuracy. In particular, CSPDarkNet53 + IRMB, CSPDarkNet53 + RCA, CSPDarkNet53 + SA, CSPDarkNet53 + IRMB + RCA and CSPDarkNet53 + IRMB+SA improve the mAP of the CSPDarkNet53 by 0.2%, 0.4%, 0.8%, 0.6%, and 1.1%, respectively. This result indicated that the three sub-modules of IRMB, RCA, and SA were effective.

4.4. Comparative Analysis

The proposed method TCSA was compared with SSD [40], YOlOv5s [41], Tea-YOLOv8s [22], AX-RetinaNet [17], and RT-DETR [42] in terms of precision, recall, and mAP, as shown in

Table 2. Performance comparison of different test models.

Model	P	R	mAP
SSD [40]	86.5	89.1	88.4
YOLOv5s [41]	87.4	82.5	88.6
Tea-YOLOv8s [22]	92.7	89.2	93.3
AX-RetinaNet [17]	91.7	90.8	93.4
RT-DETR [42]	89.6	87.8	94.2
TCSA	92.9	89.6	94.6

and Figure 7 and Figure 8, where Table 2 presents the accuracy, recall, and mAP of six different models. Figure 7 and Figure 8 primarily compare TCSA with the other five models in terms of mAP and attention heatmaps.

Table 2 shows that TCSA achieved the best precision and mAP compared to the other methods. In particular, TCSA outperformed SSD and YOLOv5 by 6.2% and 6% in mAP, respectively. These results indicated that incorporating attention mechanism plugins into deep learning frameworks effectively improved detection accuracy. The TCSA model showed significant improvements in mAP compared to Tea-YOLOv8s, AX-RetinaNet, and RT-DETR, with 1.3%, 1.2%, and 0.4%, respectively. These improvements can be attributed to TCSA’s integration of the IRMB as a backbone plugin, which effectively captures multiscale disease features. Additionally, the channel–spatial attention plugin in the neck enhances the model’s ability to refine neck regions and fuse multiscale features. These experimental results validate the effectiveness of the proposed TCSA method in tea disease detection.

Figure 7 illustrates the mAP curves for four attention mechanism-based models: TCSA, Tea-YOLOv8s, AX-RetinaNet, and RT-DETR. From Figure 7, TCSA had certain advantages in the detection of tea disease. Although RT-DETR gradually approached TCSA in average precision after 300 epochs, TCSA still showed improvement in the final results.

Figure 8 displays the attention heatmaps for TCSA, Tea-YOLOv8s, AX-RetinaNet, RT-DETR, YOLOV5s, and SSD. In the attention heatmaps, color variation indicates the degree of attention the recognition network paid to different areas of the image. The original disease images from top to bottom in Figure 8 represented scenarios such as single disease, multiscale disease, leaf occlusion, and disease in complex backgrounds. In these scenarios, our method, which integrated channel and spatial attention mechanisms, more accurately covered the target areas than other methods.

4.5. TCSA’s Performance on Different Diseases

We evaluated the performance of TCSA on various diseases using the provided dataset, as shown in

Table 3. Detection results of different diseases in TCSA.

Disease	P	R	mAP
Als	89.5	92.0	91.7
Tc	97.4	83.3	97.0
Clb	93.8	93.9	96.5
Eb	87.3	97.4	98.5
Tr	96.3	97.2	97.5
Rs	86.4	82.4	87.1

and Figure 9, where Table 3 presents the accuracy, recall rate, and mAP values of TCSA for different diseases, while Figure 9 illustrates the visual results for these diseases. As shown in Table 3, TCSA exhibited high detection accuracy for diseases such as Eb and Tr that encompass wide distribution, significant variation, and complex backgrounds. However, the mAP of Rs was relatively low at 87.1%. The mAP value exceeded the individual accuracy of 86.4% and recall rate of 82.4%, indicating that the model has excellent comprehensive performance under a variety of conditions, with good robustness and adaptability. In addition, the proposed method showed promising results in the detection of a variety of tea diseases, including Als, Tc, and Clb.

Figure 9 shows the accuracy maps generated by TCSA for the test results, including prediction boxes and scores. In Figure 9a,e, precise and accurate detection of diseases of various scales and shapes could be observed in a simple background. This demonstrated the algorithm’s ability to accurately detect scale changes. Figure 9b,e demonstrates tea disease detection under complex background and leaf occlusion conditions, accurately identifying the affected areas and disease types on the leaves. This could be attributed to the feature extraction and fusion network’s ability to accurately detect disease regions in complex backgrounds by capturing the dependencies between input feature scales. The visualization results demonstrate that the model can effectively address challenges in tea disease detection, such as varying disease scales, leaf occlusion, and complex backgrounds.

5. Conclusions

This paper proposes a TCSA network for detecting tea disease in complex backgrounds. The innovative model for tea disease detection effectively overcomes the challenges posed by complex backgrounds and variable scales. Compared to prior network models, our approach focuses more on identifying the diseased parts of tea leaf images, improving accuracy on average. In this study, TCSA achieved an outstanding mAP of 94.6%. Additionally, TCSA demonstrated excellent performance with precision and recall rates of 92.9% and 89.6%, respectively. The experimental results demonstrate that the proposed channel–spatial attention mechanism improves the feature extraction of images for tea disease diagnosis under complex backgrounds. In addition, we find that the self-attention mechanism in our method can help the backbone to effectively extract multiscale features of diseases.

Future research will focus on exploring the actual deployment and time constraints of real-time detection of tea diseases [43], strengthening the feature fusion strategy, and improving the practicability and effectiveness of the model [44]. We hope that our research will enable the accurate and timely detection of tea diseases, which will not only enhance tea production yield and quality but also contribute to achieving the agricultural circular economy and promoting sustainable agriculture.