Due to the lack of authoritative public tea datasets, the data used in this article was collected in April and May at longitude 115°8'14.54''E and latitude 32°43'47.75''N. These images were captured under natural light using a Huawei Mate60 portable device and a Sony ILME-FX30B camera, with a total of 4000 data samples collected. The pixel resolution of the image is 3024 * 4032. Among them, tea farmers and tea experts identified 43 images as red leaf spots, 213 images as algal leaf spots, 324 images as bird eye disease, 1102 images as gray wilt, 43 images as white spots, 75 images as anthrax, 1213 images as brown wilt, and 987 images as healthy leaves. Due to the limited data collected on tea defects and diseases, and the fact that the images were taken under clear weather conditions, this paper simulated adverse conditions to improve the generalization performance of the model. These simulation conditions include defocused images, partial data loss, heavy rain and snow. Data augmentation simulated conditions such as partial image loss, motion blur, early morning and dusk lighting, as well as fog, rain, snow, and wind. This method not only simulates various situations encountered in actual production, but also improves the generalization performance of the training dataset. After scaling up the original dataset by 2.5 times, a total of 10000 images were obtained. The dataset includes 10000 annotated bounding boxes (BBOX) for all defect types. Among them, 80% is the training set and 20% is the validation set. Each bounding box is manually annotated using open-source annotation tools to ensure that every defect is fully included in BBOX. Figure 1 shows a subset of the enhanced dataset.

The model in this paper adopts an improved CSPDarknet53 as the backbone network (Wang et al., 2023) for YOLOv8. It conducts down sampling on input features five times, resulting in five different scales of features, denoted as B1 to B5. The structure of the backbone network is illustrated in Figure 2F . The CSP (Cross Stage Partial) module in the original backbone network of previous versions is replaced by the C2f module. The structure of the C2f module is shown in Figure 2D , where ‘n’ represents the number of bottlenecks. The C2f module adopts gradient parallel connections, enriching the information flow of the feature extraction network while maintaining a lightweight design. The ConvModule module conducts convolutional operations on input information, followed by batch normalization, and then utilizes the SiLU activation function to obtain the output result, as shown in Figure 2C . The backbone network concludes by utilizing an improved down sampling module to pool input feature maps into fixed-size adaptive-sized outputs. Compared to the original Spatial Pyramid Pooling (SPPF) structure, the new connection layers can retain more feature information, as shown in Figure 2A .

Inspired by PANet, the original YOLOv8 incorporates a PAN-FPN structure at the neck (Wang et al., 2023). Compared to the neck structures of YOLOv5 and YOLOv7 models, YOLOv8 removes the convolutional operation after up sampling in the PAN structure, as shown in Figure 2E achieving model lightweighting while maintaining the original performance. YOLOv8 adopts a top-down and bottom-up network structure to integrate semantic information from deep and shallow features. However, this fusion is superficial. To address this, we designed a new feature fusion network based on the PAN-FPN architecture. Through the analysis of tea leaf defect images, it was determined that spatial positional information of features is not necessary in practical applications. Therefore, part of the feature information flow can be trimmed to reduce computational costs. Simultaneously, feature fusion is achieved by merging different nodes of the same feature layer, retaining more features of tea pests and diseases without increasing computational costs.

The detection part of YOLOv8 adopts a decoupled head structure, as shown in Figure 2B . This structure employs two independent branches for object classification and bounding box regression prediction, each using different loss functions. For the classification task, binary cross-entropy loss (BCELoss) is used. For the bounding box regression task, Distribution Focal Loss (DFL) and Complete Intersection over Union (CIoU) are employed. This detection structure improves detection accuracy and accelerates model convergence. YOLOv8 is an anchor-free detection model, which simplifies the specification of positive and negative samples. It also utilizes the Task-Aligned Assigner to dynamically assign samples, enhancing the detection accuracy and robustness of the model.

To focus the detection model on tea leaf defects and diseases while reducing attention on other regions, we introduce a dynamic sparse attention mechanism called Bi Former (Zhu et al., 2023) into the backbone network of the model. Bi Former utilizes adaptive querying to filter out the least relevant key-value pairs in the coarse-grained regions of the input feature map. It then efficiently identifies the key-value pairs with higher relevance and performs attention computation on them. This significantly reduces computational and storage costs, enhancing the model’s ability to perceive the input content. YOLOv8 is a convolutional neural network (CNN) model. The essence of a CNN is local processing, which limits its ability to capture relationships between global features. Compared to traditional CNN models, transformers use an attention mechanism to capture the relationships between different pieces of data, providing a global receptive field. An effective attention mechanism can build robust and powerful data-driven models, making them more flexible when handling complex, large-scale data.

The Bi Former module is designed based on a dual-stage routing attention mechanism, as shown in Figure 3 . In this block, DW Conv represents depth wise separable convolution, which reduces the number of parameters and the computational load of the model. LN stands for layer normalization, which accelerates training and improves the model’s generalization ability. MLP, or multilayer perceptron, further processes and adjusts attention weights, enhancing the model’s focus on different features. The addition symbol in Figure 3 represents the concatenation of two feature vectors.

The introduction of the Bi Former block into the backbone network in this paper serves two purposes. First, Bi Former considers the limited computational power and storage resources of mobile platforms. Second, the dynamic attention mechanism within this block enhances the model’s focus on crucial target information, thereby optimizing the model’s detection performance. To fully leverage the efficient attention mechanism of this block, we added the Bi Former block between the model backbone networks B1 and B2.

Down sampling can aggregate local information, expand the receptive field, and reduce computational costs. Conventional down sampling operations mainly involve max-pooling and stride convolution. However, pooling operations on local regions can lead to the loss of important spatial information, which is detrimental to accurate detection. To address this, we introduce down sampling operations based on the Haar (Xu et al., 2023) wavelet.

The core idea of the new down sampling operation is to use Haar wavelet transformation to reduce the spatial resolution of feature maps while preserving more information. This approach enhances the ability of semantic segmentation and reduces information uncertainty. For 2D image Haar decomposition, it can be seen as performing 1D Haar decomposition separately on all columns and all rows. Depending on the order of decomposing rows and columns, two different decomposition methods can be generated. The specific process is shown in Figure 4 .

From Figure 5 it can be seen that the module first preprocesses the input information flow in the horizontal and vertical directions by performing averaging and differencing operations on the information flow separately. Then, down sampling is performed. Next, the processed information flow undergoes diagonal direction processing, where it is averaged and differenced to obtain diagonal subbands. Each of these subbands is then down sampled. This process iteratively repeats for each subband.

Finally, an inverse transformation is applied to each subband to reconstruct the original image. These steps constitute the lossless feature encoding module primarily based on the Haar wavelet transform. Subsequently, the output information flow undergoes convolution, normalization, and activation function processing to reduce the number of channels.

YOLOv8 itself uses a simplified FPN-PANet in its neck to perform feature fusion, reducing the loss of information. The core idea of FPN is to construct a feature pyramid at different levels of the image to capture objects at different scales: by up sampling the deep feature maps to match the size of the shallow feature maps (Gong et al., 2021), and then performing an addition operation. PAN, on the other hand, employs a cascaded operation, which can retain more detailed information, thereby improving detection accuracy (Wang et al., 2019).

However, the above operations have two drawbacks: first, they do not focus on features at the same level; second, the merging process can introduce delays, leading to suboptimal merging effects. Considering that in tea plantation inspections using drones, multiple photos are taken of the same area, the edge features of a single photo are not our primary concern. At the same time, when the drone takes photos, it is approximately 50 cm away from the top of the tea trees. Each photo contains a large number of tea leaves, which implies there are many instances of defects and diseases.

To address these issues, our approach focuses on refining the feature fusion process to enhance the detection of tea leaf defects and diseases in such scenarios. By prioritizing crucial target features and optimizing the merging process, we aim to achieve more accurate and efficient detection results.

Based on the limited receptive field of CNN networks, they can only localize regions with distinctiveness. As shown in Figure 6 , Therefore, the first step is to use a multi-head attention mechanism to segment the image into patches with distinctive features. Since deep features reflect specific information about objects and require global context, a transformer encoder is used to process deep features to enhance object detection performance.

Next, under the condition of unchanged computational resources, allocating more parameters for feature fusion can be achieved by intuitively reducing the backbone layers and expanding the fusion modules. From Figure 7 , To achieve this, the backbone network is decomposed into several smaller cascaded stages, generating richer scale features. Each stage consists of a sub-backbone and a transformation module.

Performing skip connections on the same feature layer helps preserve more feature information. The Transition block utilizes 1x1 convolutions to align the channel numbers in the sampling points and uses bilinear interpolation to align the spatial sizes of features. The Focal Block, on the other hand, enlarges the convolutional kernel to expand the receptive field, thereby acquiring more feature information.

By implementing these modifications, the model can better handle complex scenes with multiple instances of tea leaf defects and diseases, improving detection accuracy and robustness.

The contributions of features from images with different resolutions are unequal, hence an additional weight is introduced for learning. Building upon Unbounded Fusion, normalization of weights is conducted to constrain the value range of each weight. Unbounded Fusion refers to integrating features from different resolutions without explicit boundaries.

To verify the positive impact of each module on the model, YOLOv8n was used as the baseline model, and ablation experiments were conducted separately on the BiFormer module, Haar module, and feature fusion module. In order to ensure the accuracy of the experimental results, the parameter settings in each individual module are the same.

At the same time, in order to ensure that the pre trained weight structure and the target model structure are the same in the experiment, all three experimental groups will undergo weight pre training before the formal experiment, and the weight pre training dataset will use the dataset from Chapter 2, which will not cause overfitting.

To demonstrate the superiority and effectiveness of the proposed improved algorithm, we conducted comparative experiments. First, we compared various models in the YOLO series: YOLOv3 (Redmon and Farhadi, 2018) nd its lightweight version YOLOv3-tiny (Adarsh et al., 2020); YOLOv4 (Bochkovskiy et al., 2020) with the novel backbone network CSPDarknet53; YOLOv5n (Xue et al., 2023), which improves accuracy using mosaic data augmentation; and YOLOv9s (Wang et al., 2024a), which introduces new structures based on YOLOv7. Also, we compared the tea detection model developed by YOLO-Tea (Xue et al., 2023), Hossain et al. (2018) and TSBA-YOLO (Lin et al., 2023), which can now be applied to the prevention and control of tea diseases and pests.

From Table 5 we can analogize the advantages and disadvantages of the model proposed in the above paper. Compared to models such as YOLOv3 and YOLOv5, the later proposed YOLOv8 and v9 have better performance, with mAP reaching over 70%. However, this is still not an ideal accuracy rate. Because these four models are only a framework and do not specifically detect the characteristics of tea pests, diseases, and defects in images. However, YOLOv10b and YOLOv11n are improvements based on YOLOv8 and YOLOv9, still retaining similar shortcomings. Therefore, subsequent research mainly focuses on targeted optimization of this drawback, such as the attention mechanism and feature fusion module proposed in this paper, which take into account the characteristics of tea damage and the features captured during drone inspections. After targeted optimization, our model achieved a precision of 92.2% and an mAP of 94.5%, far exceeding similar models.