An enhanced lightweight model for apple leaf disease detection in complex orchard environments

Automated detection of apple leaf diseases is crucial for predicting and preventing losses and for enhancing apple yields. However, in complex natural environments, factors such as light variations, shading from branches and leaves, and overlapping disease spots often result in reduced accuracy in detecting apple diseases. To address the challenges of detecting small-target diseases on apple leaves in complex backgrounds and difficulty in mobile deployment, we propose an enhanced lightweight model, ELM-YOLOv8n.To mitigate the high consumption of computational resources in real-time deployment of existing models, we integrate the Fasternet Block into the C2f of the backbone network and neck network, effectively reducing the parameter count and the computational load of the model. In order to enhance the network’s anti-interference ability in complex backgrounds and its capacity to differentiate between similar diseases, we incorporate an Efficient Multi-Scale Attention (EMA) within the deep structure of the network for in-depth feature extraction. Additionally, we design a detail-enhanced shared convolutional scaling detection head (DESCS-DH) to enable the model to effectively capture edge information of diseases and address issues such as poor performance in object detection across different scales. Finally, we employ the NWD loss function to replace the CIoU loss function, allowing the model to locate and identify small targets more accurately and further enhance its robustness, thereby facilitating rapid and precise identification of apple leaf diseases. Experimental results demonstrate ELM-YOLOv8n’s effectiveness, achieving 94.0% of F1 value and 96.7% of mAP50 value—a significant improvement over YOLOv8n. Furthermore, the parameter count and computational load are reduced by 44.8% and 39.5%, respectively. The ELM-YOLOv8n model is better suited for deployment on mobile devices while maintaining high accuracy.

1 Introduction

As one of the most widely cultivated fruits globally, the annual production of apples significantly influences the agricultural economy and the global food supply chain (Wani and Mishra, 2022). As the largest producer of apples worldwide, China contributes 40.5 million tons annually, representing 46.7% of the global production (Liu et al., 2024b). However, the yield and quality of apples are affected by a variety of diseases, such as Rust, Grey Spot, Powdery Mildew, and Scab, which not only reduce the quality of fruits, but also directly lead to a decrease in yield. Therefore, accurate and timely identification and detection of leaf diseases is crucial during apple cultivation. This can assist farmers in more effectively controlling the spread of diseases, thereby maintaining apple quality and production levels, and ultimately achieving both economic and environmental benefits. Traditional identification of crop diseases primarily depends on the expertise of agricultural specialists and visual assessment. This method is inefficient and limited by the expert’s knowledge and experience, which is prone to misjudgment and omission. With the development of computer technology, machine learning has started to find application in the recognition of crop diseases (Ramesh et al., 2018). However, traditional feature extraction algorithms usually need to rely on manual design and domain expertise to select and construct features. Compared to traditional machine learning methods, deep learning methods can automatically learn and extract features, reducing the workload of manual feature engineering and exhibiting better performance when dealing with large-scale data. In 2020, Li et al. (2020) introduced a finely-tuned GoogLeNet model, which was fine-tuned to improve over the state-of-the-art method by 6.22%.In 2021, Soujanya and Jabez (2021) proposed a technique based on improved AlexNet that can successfully classify 38 different classes of healthy and diseased plants, and the best accuracy of this method is 96.5%.In 2022, Wei et al. (2022) proposed a multi-scale feature fusion based network model for accurate identification and classification of crop pests and diseases, the classification accuracy reached 98.2%.In 2023, Islam et al. (2023) used ResNet50 migration learning model as the core for distinguishing between healthy and infected leaves and classifying current disease types, providing an even higher accuracy of 98.98%. In 2024, Faisal et al. (2024) proposed a method based on deep CNN architecture to recognize and classify cotton weeds efficiently and the proposed model achieved 98.3% accuracy which is better than other models. Although these studies have achieved satisfactory accuracy rates when dealing with the task of classifying a single disease image with a simple background, however, the performance may still be insufficient in real-world applications when faced with more complex and varied scenarios, as well as in mobile deployments.

Faced with real-time detection in complex environments, object detection algorithms (Ren et al., 2016; Liu et al., 2016; Redmon, 2016) show greater advantages. Sun et al. (2021) proposed a lightweight CNN model, MEAN-SSD, which can be deployed on mobile devices for real-time detection of apple leaf diseases. Khan et al. (2022) proposed a two-stage real-time detection system for apple leaf diseases based on Xception and Faster-RCNN-based two-stage real-time apple disease detection system, which achieved an overall classification accuracy of 88%. Zhang et al. (2023b) designed BCTNet network for accurate apple leaf disease detection, which solved the problems of unconstrained environmental factor interference and low detection accuracy caused by significant changes in the target scale of apple leaf disease detection. Lv and Su (2024) proposed YOLOV5-CBAM-C3TR for apple leaf disease detection, which showed strong recognition ability in identifying similar diseases, and is expected to promote the further development of disease detection technology. Chang and Lai (2024) integrated a lightweight convolutional neural network architecture, RegNetY-400MF, with a transfer learning technique, which not only improves the accuracy of potato leaf disease detection, but also reduces the computational and storage requirements. Liu et al. (2024a) proposed MCDCNet to extract more reliable apple leaf disease features with various scales and geometries, which effectively improved the discriminative ability of the network. However, the reduction of model parameters and computational load may lead to the decrease of accuracy, how to strike a balance between lightness and accuracy is a hot research topic nowadays (Yan et al., 2024).

In summary, although the crop disease recognition method based on Convolutional Neural Networks (CNN) solves the problem of inefficiency of traditional machine vision recognition, there are still problems such as high model complexity and low accuracy of small target disease recognition under complex background. To address these challenges, this study introduces a lightweight detection model for identifying apple leaf diseases against complex backgrounds. First, the Fasternet module is integrated to decrease the number of parameters and computational load of the model; second, an Efficient Multi-Scale Attention (EMA) is incorporated to improve the feature extraction ability of the network for small target diseases as well as the differentiation ability for similar diseases; then, the DESCS-DH detection head is designed to optimize the model’s effectiveness in the face of multi-scale target detection. Finally, the NWD loss function is employed to improve the model’s precision in localizing and identifying small targets, thereby enhancing the model’s robustness and facilitating rapid and accurate identification of apple leaf diseases.

2 Related works

With the development of artificial intelligence technology, machine learning and deep learning have been widely used in the field of automatic crop disease detection. The following is a brief overview of these two methods.

2.1 Machine learning methods

This approach first requires the extraction of disease features from collected plant images, which are subsequently used to train machine learning models that typically require more domain knowledge and human intervention to select and optimize features. In 2022, Harakannanavar et al. (2022) developed an algorithm based on machine learning and image processing to automatically detect tomato leaf disease. SVM, KNN, and CNN were used to classify features, and the accuracy of the three methods used was 88%, 97%, and 99.6%, respectively. In 2023, Gaikwad and Musande (Gaikwad and Musande, 2023) developed an advanced crop disease prediction technique using cetalatran-optimization algorithm, deep KNN, and relief algorithm, the accuracy reached 91.879%. Ahmed and Yadav (2023) developed a crop disease prediction model using back propagation ANN, SVM, GLCM, and k-mean algorithm, the accuracy reached 99%.

2.2 Deep learning methods

Deep learning methods, such as convolutional neural networks (CNN), can learn features directly from the original image without the need for a complex feature extraction process. At present, the research on deep learning in the field of plant disease diagnosis is developing rapidly, which provides strong technical support for agricultural production. Many researchers use famous CNN architectures such as AlexNet, VGG, GoogLeNet, and ResNet for disease classification. Anim-Ayeko et al. (2023) proposed a ResNet9 model that detects the blight disease in both potato and tomato leaf images for farmers to leverage, the model achieved 99.25% accuracy. Reddy et al. (2023) proposed a customized PDICNet model for crop disease identification and classification, with an accuracy and F1-score of 99.73% and 99.78%, respectively, for PlantVillage dataset. Current models based on disease classification have reached a high accuracy.

To further localize where the disease is located, two-stage detection algorithms and single-stage detection algorithms have been applied to disease detection, such as Faster R-CNN, YOLO, and SSD. We chose advanced object detection models for apple leaf disease detection and analyzed them in comparison with our model in comprehensive aspects. As shown in Table 1, for dataset selection, we synthesized the two most commonly used publicly available apple disease datasets to ensure diversity of data sources. In terms of disease types, we chose a wider range of disease types, selecting six of the most common ones. In terms of research methodology, we not only pay attention to the improvement of model performance, but also focus on the resource consumption of the model. In addition to improving the traditional feature extraction module, we also proposed a Detail-Enhanced Shared Convolutional Scaling Detection Head, so that the model can more accurately capture the subtle features of the disease when facing multi-scale targets. The model we developed not only achieves excellent accuracy, with a mAP50 of 96.7%, but also optimizes the number of model parameters and the amount of computation, effectively alleviating the problem of resource consumption. In terms of the adaptability of application scenarios, our model is able to cope with complex field environments, showing good practicality and flexibility.

Table 1

Table 1. Comparison of object detection models for apple leaf disease detection.

3 Materials and methods

3.1 Dataset construction

The experimental datasets were sourced from the publicly available datasets plant-pathology-2021-fgvc8 and AppleLeaf9.Both datasets contain a large number of high-quality images of apple leaf diseases, from which six leaf diseases were selected for labeling, namely Rust, Scab, Grey Spot, Frog Eye Leaf Spot, Powdery Mildew, and Alternaria Blotch, with more than 90% of the images collected in the orchard environments. These six diseases are prevalent on apple leaves, resulting in significant detrimental effects on agricultural productivity, and they are representative in terms of symptom presentation, covering different lesion types such as spotting, powdering, and rusting. There are differences in the difficulty of recognizing different diseases. For instance, rust is typically more easily recognizable by the model because of its distinctive color and shape feature, whereas Alternaria Blotch and Frog Eye Leaf Spot impose greater demands on the model’s discriminatory capabilities due to their more similar lesion characteristics.

A total of 2,203 images were labeled using an online platform, Make Sense. The small target box is used to label Rust, Grey Spot, Frog Eye Leaf Spot, and Alternaria Blotch. For Scab and Powdery Mildew, which develop throughout the leaf, the entire leaf is labeled. The dataset was augmented through data enhancement techniques. To prevent both the original and enhanced images from appearing simultaneously in the training and validation sets, the original images were initially divided into training, validation, and test sets in a ratio of about 8:1:1.Our model requires a large amount of data to capture the subtle differences between different diseases, and this ratio ensures that the model has enough data for training and the right amount of data for validation and testing. Subsequently, the images and labels were augmented using five techniques: horizontal flipping, vertical flipping, translation, contrast adjustment, and brightness adjustment, the new dataset obtained is named Apple-Leaf1.An example of the enhancement methods is illustrated in Figure 1. The numbers and sources of the images are presented in Table 2.

Figure 1

Figure 1. Data enhancement methods.

Table 2

Table 2. Number and source of images in the Apple-Leaf1 dataset.

3.2 Methodology

3.2.1 Fasternet module

In order to design lighter and faster networks, many researchers have generally focused on reducing the total amount of floating-point operations (FLOPs) (Menghani, 2023; Zeng et al., 2022; Zhong et al., 2022). However, merely reducing FLOPs does not invariably result in the desired reduction in delay. The underlying reason lies in the inefficiency of the actual floating-point operations per second (FLOPS) performed. Chen et al. (2023) designed a new convolution (PConv), which is able to significantly reduce unnecessary computations and memory accesses while extracting spatial features through a subtle design, thus significantly improving the computational efficiency. Based on this, the FasterNet family was further developed, as shown in Figure 2, which is a new set of neural network architectures.

Figure 2

Figure 2. Fasternet Block. The "*" represents Convolution.

In this study, we integrate the Faster Block into the backbone and neck to obtain C2f-Faster, which reduces the number of parameters and computational load of the model. The Faster Block is designed to fully incorporate the high efficiency of PConv.The FLOPs of PConv, as shown in Equation 1, is 1/16 of the conventional convolution.

$\begin{array}{ll}h\times w\times k^2\times c_{\text{p}}^2& (1)\end{array}$

And the memory access of PConv, as shown in Equation 2, is 1/4 of the conventional convolution.

$\begin{array}{ll}h\times w\times 2c_p+k^2\times c_p^2\approx h\times w\times 2c_p& (2)\end{array}$

The Faster Block achieves a significant reduction in computational complexity and memory access by introducing PConv, which applies regular convolution operations to a portion of the input channel while leaving the remaining channels unchanged in their original state. Although only a portion of the input channel is computed, the preserved channel remains useful in subsequent PWConv layers, allowing feature information to flow through all channels. This design enables the Faster Block to operate rapidly on various hardware platforms, including mobile devices, while maintaining high detection accuracy.

3.2.2 EMA module

In computer vision tasks, channel or spatial attention mechanisms effectively extract local features of an object (Sun et al., 2022; Wang et al., 2024b; Zheng et al., 2023). Ouyang et al. (2023) proposed the Efficient Multi-Scale Attention, which pursues a balance between computational efficiency and information retention by adopting an innovative approach: reconstructing some channels into batch dimensions and subdividing the channel dimensions into multiple sub-feature groups. This ensures a uniform distribution of spatial semantic features within each sub-feature group, thereby optimizing feature expression and processing without adding computational burden. As illustrated in Figure 3, the EMA module first divides the input into multiple grouped feature maps, which are then processed through three parallel path branches: two perform one-dimensional global pooling, and the third conducts feature extraction via 3x3 convolution. Ultimately, the output feature maps within each group are aggregated by computing the two generated spatial attention weights and then processed using a Sigmoid function to capture pairwise pixel relationships and emphasize the global contextual information of all pixels.

Figure 3

Figure 3. EMA module.

The onset sites of apple leaf diseases typically appear as small spots on the leaves, particularly during the early stages of infection. However, these small lesions are difficult to detect due to the complex background. Furthermore, the initial symptoms of several apple leaf diseases, such as Alternaria Blotch, Grey Spot, and Frog Eye Leaf Spot, are remarkably similar, making them challenging to distinguish. To address the issue of insufficient feature extraction for small target disease against a complex background, the EMA mechanism is integrated into the deep layers of the network’s backbone, as illustrated in Figure 4. Feature extraction is conducted in depth following Pconv and two 1×1 convolution operations, enabling the model to focus more on disease details and thereby enhancing recognition accuracy.

Figure 4

Figure 4. Architecture for Fasternet-EMA.

3.2.3 DESCS-DH (detail-enhanced shared convolutional scaling detection head)

Although YOLOv8 performs well on many tasks, there is still room for improvement in the design of the detection head. The first is the large number of parameters, the original detection head employs one 1×1 and two 3×3 convolutions for feature extraction before predicting the object class and positional offset within each bounding box, which inevitably results in a considerable parameter increase; second is that the use of a normal convolution does not capture the detailed information of the image very well during feature extraction, which may lead to important information loss; and lastly, the reliance on predefined anchors may cause the model to perform poorly in detecting targets of different scales and proportions, as the anchors may not be well adapted to the shapes and sizes of all objects. To address these challenges, we design a detail-enhanced shared convolutional scaling detection head named DESCS-DH.

DESCS-DH is an efficient and lightweight detection head, which can maintain high accuracy with reduced number of parameters and computation. As shown in Figure 5, firstly, DESCS-DH significantly reduces the number of parameters by fusing P3, P4, and P5 feature maps at different scales through a shared convolutional layer for feature interaction and parameter sharing. Secondly, it uses detail-enhanced convolution to improve the ability of the detection head to capture image details. DEConv significantly enhances the model’s representation and generalization performance by incorporating a priori information into the ordinary convolutional layers. In addition, through the application of reparameterization techniques, DEConv can be equivalently converted to ordinary convolution, a process that does not require the addition of extra parameters or computational cost, thus effectively improving performance while keeping the model lightweight. Finally, DESCS-DH can adaptively modify its internal parameters based on the dimensions of the input image and feature map, so that when facing multi-scale targets, each detector head can obtain the optimal working parameters, which enables it to maintain stable performance in a variety of complex environments, thereby enhancing the robustness of the entire model and its detection efficiency.

Figure 5

Figure 5. The structure of DESCS-DH. (A) DESCS-DH (B) DEConv.

3.2.4 NWD loss function

In the field of object detection, Intersection over Union (IoU) has long been a standard metric for measuring the accuracy of bounding boxes (Peng and Yu, 2021; Zhang et al., 2022; Tian et al., 2022). However, for the detection of tiny objects, the performance of IoU is not satisfactory, which is mainly due to the small overlapping area of the bounding box of tiny objects, resulting in a low value of IoU, which is difficult to accurately reflect the actual detection accuracy. To solve this problem, Wang et al. (2021) proposed a novel metric based on the Wasserstein distance, aiming to evaluate the similarity of the bounding boxes of tiny objects more effectively.

Specifically, the method first models the shape and position information of the bounding box as a two dimensional Gaussian distribution. This modeling approach captures the characteristics of the bounding box in more detail, especially for objects with irregular shapes or small sizes. Next, the Normalized Wasserstein Distance (NWD) is introduced to quantify the similarity between these Gaussian distributions. An additional significant benefit of the NWD is its insensitivity to object scale, enabling more precise localization for tiny objects. In addition, the application of NWD is not limited to the label assignment stage; it can also replace the IoU in the non-maximal suppression (NMS) process and be used in the regression loss function, thereby improving the performance of tiny object detection in general. The computational procedure is as follows, after modeling the bounding box as a two dimensional (2D) Gaussian distribution, the distribution distance is calculated using the Wasserstein distance metric. For two dimensional Gaussian distribution, u1 and u2 as shown in Equation 3:

$\begin{array}{ll}{\mu }_1=\mathcal{N}({\text{m}}_1,{\text{Σ}}_1),{\mu }_2=\mathcal{N}({\text{m}}_2,{\text{Σ}}_2)& (3)\end{array}$

Define the distance between them as Equation 4:

$\begin{array}{ll}{W}_2^2({\mu }_1,{\mu }_2)=\Vert m_1-m_2{\Vert }_2^2+Tr({\text{Σ}}_1+{\text{Σ}}_2-2{({\text{Σ}}_2^{1/2}{\text{Σ}}_1{\text{Σ}}_2^{1/2})}^{1/2})& (4)\end{array}$

The simplified formula is shown in Equation 5:

$\begin{array}{ll}{W}_2^2({\mu }_1,{\mu }_2)=\Vert m_1-m_2{\Vert }_2^2+\Vert {\text{Σ}}_1^{1/2}-{\text{Σ}}_2^{1/2}\Vert {}_F^2& (5)\end{array}$

Furthermore, for Gaussian distributions $N_a$ and $N_b$, which are derived from bounding boxes $\text{A }=(\text{c}{\text{x}}_{\text{a}},\text{ c}{\text{y}}_{\text{a}},{\text{w}}_{\text{a}},{\text{h}}_{\text{a}})$ and $\text{B }=(\text{c}{\text{x}}_{\text{b}},\text{ c}{\text{y}}_{\text{b}},{\text{w}}_{\text{b}},{\text{h}}_{\text{b}})$, Equation 5 can be simplified to Equation 6:

$\begin{array}{ll}{W}_2^2({\mathcal{N}}_a,{\mathcal{N}}_b)=\Vert {({[c{x}_a,c{y}_a,\frac{w_a}{2},\frac{h_a}{2}]}^{\text{T}},{[c{x}_b,c{y}_b,\frac{w_b}{2}.\frac{h_b}{2}]}^{\text{T}})\Vert }_2^2& (6)\end{array}$

However, the numerical range of $W_2^2(N_a,N_b)$ is not suitable for direct use as a similarity metric. Consequently, its exponential form is adopted for normalization, yielding a new metric termed the Normalized Wasserstein Distance (NWD), as shown in Equation 7:

$\begin{array}{ll}NWD({\mathcal{N}}_a,{\mathcal{N}}_b)=\text{exp }(-\frac{\sqrt{W_2^2({\mathcal{N}}_a, {\mathcal{N}}_b)}}{C})& (7)\end{array}$

In the detection and localization of small target diseases, the limited pixel occupation in the image results in scarce semantic information, making it challenging for the network to extract features that are sufficiently effective for precise target categorization and localization. The NWD loss function exhibits a unique sensitivity to both the size of the target and the distances between targets. Traditional loss functions may not fully consider these factors. For apple leaf disease detection, accurately locating and distinguishing spots of different sizes and location relationships is crucial for enhancing detection accuracy. The NWD loss function facilitates the model’s focus on features pertinent to disease detection during training, owing to its sensitivity to target size and inter-target distances.

3.2.5 ELM-YOLOv8n network architecture

YOLOv8, an advanced object detection algorithm, effectively strikes a balance between detection accuracy and speed (Wan et al., 2024; Zhu et al., 2024). Consequently, in this study, we selected and enhanced the YOLOv8n model to develop a new network architecture termed ELM-YOLOv8n.The YOLOv8 architecture comprises four components: Input, Backbone, Neck, and Head. The Backbone extracts features from the input image data using CSPDarknet53 to convert raw pixels into high-level semantic features. The Neck serves as the middle layer of the network structure and fuses the features of different levels through techniques such as Feature Pyramid Network (FPN).The Head is the output layer of the network, which predicts the information of the target’s category, location, and confidence level based on the fused features. As shown in Figure 6, the enhanced lightweight model ELM-YOLOv8n is designed in this research to realize the model to reduce the consumption of computational resources while maintaining high detection performance. The C2f module in the original YOLOv8 consumes a significant portion of the network’s parameters. By substituting the Bottleneck with the Fasternet Block, we integrate the C2f-Faster structure, which effectively diminishes the parameter count and model complexity. The EMA attention is incorporated into the deepest layer of the Fasternet Block in Backbone, and after Pconv and two 1×1 convolutional operations, the deep subtle features are extracted. The Head employs the DESCS-DH, addressing the issues of high parameter count and inadequate detail capture in the original detector head, as well as inconsistencies in performance across varying target sizes. Lastly, the NWD loss function for small target detection is utilized in place of the CIOU loss function to enhance the model’s accuracy in localizing small targets.

Figure 6

Figure 6. ELM-YOLOv8n overall framework and its component modules (A) ELM-YOLOv8n. (B) Conv. (C) SPPF. (D) C2f-Faster. (E) C2f-Faster-EMA.

3.3 Experimental environment

3.3.1 Experimental platform

The experiment was conducted using a 64-bit Windows 11 operating system, with the PyTorch framework, PyCharm IDE, and Python programming language employed for model training. To enhance data diversity and mitigate overfitting, the mosaic augmentation technique was applied during the training phase. The detailed experimental environment and some experimental parameters are presented in Tables 3, 4, respectively.

Table 3

Table 3. Experimental environment.

Table 4

Table 4. Experimental parameters.

3.3.2 Evaluation metrics

The model is evaluated on two dimensions: performance and complexity. Performance metrics include precision (P), which measures the proportion of positive samples detected by the model that are actually positive; recall (R), which refers to the proportion of samples that are actually in the positive category that are correctly predicted by the model to be in the positive category; F1 score, which is the harmonic mean of precision and recall; average precision (AP), which is a combination of precision and recall and is obtained by plotting a precision-recall curve and calculating its area; and mean average precision in multiple categories(mAP), which calculates the AP for each category and averages the values. The specific calculation formulas are shown in Equations 8–12:

$\begin{array}{ll}\text{P}=\frac{\text{TP}}{\text{TP}+\text{FP}}\times 100\%& (8)\end{array}$

$\begin{array}{ll}\text{R}=\frac{\text{TP}}{\text{TP}+\text{FN}}\times 100\%& (9)\end{array}$

$\begin{array}{ll}F1=\frac{2\times P\times R}{P+R}\times 100\%& (10)\end{array}$

$\begin{array}{ll}\text{AP}={\int }_0^1\text{P }(\text{R})\text{ dR}\times 100\%& (11)\end{array}$

$\begin{array}{ll}\text{mAP}=\frac{1}{N}{\displaystyle \sum _{\text{k}=1}^{\text{k}=\text{N}}}{\text{AP}}_{\text{k}}\times 100\%& (12)\end{array}$

In complexity evaluation, three main indicators are considered: the number of model parameters(Params), Giga Floating-point Operations Per Second (GFLOPs), and the model size. The number of parameters of the model is the total number of all trainable parameters in the model, which are optimized during training to reduce prediction errors.The parameter count serves as a crucial indicator of a model’s complexity. GFLOPs is used to measure the amount of model computation. Model size denotes the memory space required for storing the model on disk and aids in evaluating its deployment viability in environments with limited resources.

4 Results and analyses

4.1 Effect of incorporating different attention mechanisms into the backbone C2f on model performance

Incorporating different attention mechanisms into the deepest layer of the Fasternet Block in the backbone network, the attention mechanisms together with the lightweight module resulted in a reduction in the number of parameters and computation of the whole model. As shown in Table 5, the CF-LSKA has the highest precision of 94.6%. In the F1 score, a composite metric, both CF-TripleAttention and CF-EMA achieve 92.5%, indicating that they strike a good balance between precision and recall. On the key metrics of mAP50 and mAP50-95, CF-EMA performs well with 96.3% and 66.5%, respectively. Finally, CF-EMA shows excellent efficiency in terms of parameter count and computation, with its parameter count of only 2.65M and its GFLOPs remaining at 7.1, which means that it maintains high performance with high computational efficiency and low resource consumption.

Table 5

Table 5. Effect of different attention mechanisms on model performance.

4.2 Effect of different detection heads on model performance

Among the four different detection heads, with the exception of the Aux detection head (Wang et al., 2023), the other three detection heads reduce the number of parameters and computational complexity of the model. Specifically, LADH (Zhang et al., 2023a) has the lowest computational complexity, but this is accompanied by lower accuracy. As presented in Table 6, LADH has relatively low F1 scores and mAP50 values. The Aux detection head has the highest accuracy. However, on the whole, the DESCS-DH has relatively high F1 score and mAP50 value, and also has a low number of parameters and computational complexity, thus achieving a balance between performance and computational resources.

Table 6

Table 6. Effect of different detection heads on model performance.

4.3 Performance comparison of different loss functions

Model training and validation were performed using different loss functions. As presented in Table 7, the NWD loss function yields the highest precision, F1, and mAP50, with respective values of 94.1%, 93.5%, and 96.6%. As shown in Figure 7, all five loss functions significantly decreased the model’s boundary regression loss throughout the network’s training process. However, the DIOU, CIOU and EIOU resulted in relatively high loss value, whereas the NWD loss function achieved the lowest loss value.

Table 7

Table 7. Effect of different loss functions on model performance.

Figure 7

Figure 7. Change in loss value.

4.4 Ablation experiments

As shown in Table 8, the F1, mAP50, and mAP50-95 of YOLOv8n on the test set are 92.3%, 95.5%, and 66.2%, respectively. Following the integration of Fasternet Block, the model’s parameter count and computational complexity were significantly reduced, accompanied by a minor decline in the F1 score and a slight enhancement in the mAP50 metric. After incorporating the EMA mechanism, only a slight increase in model parameters and computational effort was exchanged for a significant improvement in F1 score and mAP50-95 value, by 1.4% and 1.2%, respectively. The adoption of the DESCS-DH detection head further significantly reduced the model’s complexity, yielding a 0.1% increase in mAP50 value and a 0.2% reduction in the F1 score. With the introduction of the loss function NWD, the parameters and computational effort of the model did not change, but the F1, mAP50, and mAP50-95 improved by 0.9%, 0.2%, and 0.6%, respectively. Consequently, the model’s F1 score and mAP50 value were enhanced by 1.7% and 1.2%, respectively, compared to the baseline YOLOv8n model. Additionally, the model’s parameter count and computational effort were reduced by 44.8% and 39.5%, respectively.

Table 8

Table 8. Results of ablation experiments.

Figure 8 illustrates that the ELM-YOLOv8n model exhibits superior accuracy in identifying various apple diseases compared to the original YOLOv8n model, with the performance improvement being particularly pronounced for the detection of small target diseases. For example, the mAP50 indexes for detecting Grey Spot, Alternaria Blotch, and Frog Eye Leaf Spot, which are small target diseases, were improved by 2.0%, 2.2%, and 1.0%, respectively.

Figure 8

Figure 8. MAP50 values of YOLOv8n and ELM-YOLOv8n in the detection of six diseases.

4.5 Comparison with other models

Table 9 presents the performance comparison of the seven object detection algorithms on the test set. ELM-YOLOv8n achieves the highest F1 and mAP50 scores of 94.0% and 96.7%, respectively, where mAP50 compares favorably with Faster R-CNN, RT-DETR-l, YOLOv5n, YOLOv7-tiny, YOLOv8-Mobilenetv4, and YOLOv10n by 24.2%, 1.8%, 2.1%, 0.9%, 3.3%, and 0.2%, respectively. Figure 9 depicts the changes in mAP50 for various models throughout the training process. With respect to model complexity, ELM-YOLOv8n has the fewest parameters, totaling only 1.66M. Regarding computational and model size, as illustrated in Figure 10, Faster R-CNN exhibits the highest computation and largest model size, at 948.2 GFLOPs and 108.0 MB, respectively. In contrast, YOLOv5n has the smallest computation and model size, followed by the improved model, but the improved model’s F1, mAP50, and mAP50-95 values are 2.5%, 2.1%, and 3.1% higher than YOLOv5n, respectively.

Table 9

Table 9. Performance comparison of different network models on test sets.

Figure 9

Figure 9. Changes in mAP50 during training of different models.

Figure 10

Figure 10. Comparison of computation and model size for different models.

4.6 Generalization experiment

To evaluate the model’s adaptability to new data and its performance in various environments and scenarios, we conducted generalization experiment. A collection of 823 apple disease images from different scenarios were selected to be collected under different light, weather and geographic locations, with 755 images from field scenarios and 68 images from simple scenarios to simulate a variety of situations in the real world, thus testing the stability of the model under these conditions. Images were selected to cover a wide range of disease types including Rust, Scab, Grey Spot, Frog Eye Leaf Spot, Powdery Mildew, and Alternaria Blotch, naming this dataset Apple-leaf2.The example of the dataset is illustrated in Figure 11.

Figure 11

Figure 11. Example of Apple-leaf2 dataset.

Experimental results show that the model can work efficiently in simple scenarios as well as remain robust in complex scenarios. As shown in Table 10, the ELM-YOLOv8n model exhibits an enhanced F1 score of 0.8% and an mAP50-95 value improvement of 1.1%, relative to the YOLOv8n. Concurrently, the parameter count and computational complexity are diminished by 44.8% and 39.5%, respectively. These results suggest that the improved model has strong detection capabilities for leaf diseases in multifactorial complex scenarios and is more suited for deployment on mobile devices.

Table 10

Table 10. YOLOv8n and ELM-YOLOv8n detection results on Apple-leaf2 dataset.

To intuitively illustrate the interpretability of the model’s decisions, we utilize XGrad-CAM to visualize the model’s predictions. This method involves computing the gradient of the feature map in the convolutional neural network relative to the target category, combining these gradients with the feature map values to obtain importance scores, and normalizing the result to create a heat map that highlights the critical regions in the image for decision-making. Figure 12 displays the heat maps comparison between YOLOv8n and ELM-YOLOv8n on selected test sets. The heat maps generated by ELM-YOLOv8n reveal that the predicted targets exhibit darker colors, signifying that the model allocates greater attention to regions of interest within the target image during the prediction.

Figure 12

Figure 12. Comparison of YOLOv8n and ELM-YOLOv8n heat maps on selected test sets.

5 Discussion

Current apple disease detection algorithms confront various challenges, including large model parameters, high computational complexity, and long inference times (Ma et al., 2024). To effectively address these issues and achieve accurate detection of apple leaf diseases, this study introduces the Fasternet Block and proposes a detail-enhanced shared convolutional scaling detection head (DESCS-DH) for constructing a lightweight, real-time object detection model. Incorporating the Fasternet Block into C2f reduced the model’s parameters and computational requirements by 23.5% and 22.2%, respectively. The F1 score slightly decreased by 0.4%, while the mAP50 increased by 0.4%. This can be attributed to PConv in Fasternet, which convolves only a subset of input channels, significantly reducing model complexity and storage space needs. This allows the Fasternet Block to effectively manage the number of model parameters with minimal performance degradation. Utilizing DESCS-DH as the detection head for YOLOV8, the model’s mAP50 is enhanced by 0.8% while simultaneously decreasing model complexity. DESCS-DH is an efficient and lightweight detection head that boasts several advantages, including the use of shared convolutions for parameter sharing, which drastically decreases the number of parameters. In the task of apple leaf disease detection, critical feature information of the target, such as disease edges and textures, is often found in image details. Standard convolution operations may overlook these subtle yet vital details; however, DEConv within DESCS-DH effectively captures them. When combined with re-parameterization techniques, DEConv optimizes the convolution operation without increasing the model’s computational load, thereby enhancing the model’s responsiveness to input data. In real-world disease detection scenarios, the target sizes vary significantly. DESCS-DH can adaptively modify its internal parameters, ensuring that the detection head maintains stable performance across a range of target sizes.

In the recognition of small target diseases on apple leaves within complex backgrounds, the primary challenges are the small size of target objects, their subtle features, and their susceptibility to interference from background noise, lighting variations, leaf occlusion, and similar textures. These challenges compound the difficulty of detection, thereby diminishing recognition accuracy (Shao et al., 2024; Wang et al., 2024a). Incorporating the EMA mechanism into YOLOv8n enhances the mAP50 by 0.8%. The EMA mechanism constructs local cross-channel interactions within each parallel sub-network without reducing channel dimensions. Through a cross-space learning approach, it fuses the output feature maps of the two parallel sub-networks and generates better pixel-level attention for high-level feature maps. Thus, the model can focus on the disease area and reduce the influence of disturbing factors, even in the presence of light variations or background noise. Spatial attention weights, processed by the sigmoid function, capture pairwise pixel relationships, facilitating the model’s comprehension of structural information within the diseased area and enabling the recognition of minute disease features through pixel interactions, even when such features are not readily apparent. In the detection and localization of small target lesions, they occupy fewer pixels in the image, which makes it difficult for the model to accurately determine the category and location of the target, while the NWD (Normalized Wasserstein Distance) loss function is sensitive to the size and distance of the target. Utilizing the NWD as a loss function for YOLOV8, the model’s F1 score is enhanced by 1.2%, and the mAP50 value is incremented by 1.1%.

Apple orchards usually have a large area, and traditional artificial disease detection methods are not only extremely inefficient, but also difficult to achieve comprehensive and timely monitoring. The enhanced lightweight model proposed in this study has significant features such as high precision, low computational complexity and small model size, which makes it very suitable for deployment on mobile devices with limited resources. In the next research, we plan to deploy the model on the UAV. By integrating edge computing devices on the UAV, the model can process the collected image data in real time during UAV flights, timely and accurately identify diseases in the orchard, and provide support for the disease prevention and control of the orchard. The UAV can automatically fly according to the preset route and complete the scanning of the entire orchard in a short time, which greatly improves the detection efficiency.

The ideal UAV platform should have the following characteristics: sufficient computing power(GPU over 2GB of VRAM), a high-resolution camera, equipped with a Global Positioning System (GPS), automatic flight control, wireless communication systems, multiple sensors, and other functional components, preferably a multi-rotor UAV (Chen et al., 2021). However, the accuracy of the model is affected by the distance of the drone’s image capture and the flight speed. Based on theoretical analysis, in order to ensure a certain field of view and obtain a clear image of the leaf, the appropriate flight altitude should be maintained 3-5 meters (Hou et al., 2023), which can meet the needs of most disease detection. With the improvement of camera resolution, the height can be increased to more than 5 meters. The appropriate flight speed should be set 1.5-3 m/s. By adjusting the shooting parameters, most common diseases can be accurately recognized. As the flight speed increases, the recognition accuracy of the model will decline, and there may be a small amount of missed detection for some extremely subtle disease features, such as the early mild symptoms of apple powdery mildew. In addition, the present model has some limitations in dealing with the presence of multiple diseases in a leaf, because the training data primarily target the case of a single disease. Therefore, the model still requires further optimization and improvement for the identification of multiple diseases. We will work on these challenges in our future work to enhance the comprehensive performance and application scope of the model.

6 Conclusions

To achieve real-time detection of apple leaf diseases in complex environments, this study introduces an enhanced lightweight model, ELM-YOLOv8n.In order to overcome the challenge of inadequate feature extraction from small targets in complex backgrounds, we incorporate the Efficient Multi-Scale Attention (EMA), which captures both channel and spatial information, enhancing feature representation without a significant increase in parameters or computational cost. To decrease the parameter count and computational load, the Fasternet Block is integrated into the C2f module of the backbone and neck, substantially reducing the model’s complexity. In order to make the network more lightweight and better extract the edge information of the disease, we design the Detail Enhanced Shared Convolutional Scaling Detection Head (DESCS-DH), which can maintain the high accuracy with fewer parameters and computation. The detection head can adaptively modify its internal parameters in response to the input feature map size, ensuring optimal operating parameters for multi-scale targets, thereby enhancing resilience to environmental changes and interference. Lastly, utilizing the NWD loss function enhances the model’s precision in localizing and identifying small targets, which further enhances the model’s recognition accuracy.

The experimental results indicate that the ELM-YOLOv8n algorithm performs well in terms of computational resource consumption while maintaining excellent detection accuracy, which fully meets the demand for real-time processing. Compared to other current models, this approach not only significantly improves the detection accuracy, but also drastically reduces the demand for computing platform resources, making it more feasible to be deployed on resource-constrained devices. Future research will concentrate on enhancing the model’s robustness across various environmental conditions and on optimizing its capability to detect and process a broader range of crop diseases. The optimized model is poised for successful application in resource-limited embedded detection systems, with the algorithm set to undergo further refinement to guarantee its efficiency and reliability in practical settings.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

Author contributions

GW: Conceptualization, Funding acquisition, Supervision, Writing – original draft, Writing – review & editing. WS: Conceptualization, Data curation, Investigation, Methodology, Software, Writing – original draft, Writing – review & editing. FX: Methodology, Validation, Writing – review & editing. YG: Software, Validation, Writing – review & editing. YH: Data curation, Validation, Writing – review & editing. QL: Funding acquisition, Writing – review & editing.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. This work was supported by the Shandong Province Science and Technology-based Small and Medium-sized Enterprises Innovation Capacity Enhancement Project[2024TSGC0158].

Acknowledgments

The authors are very grateful to the editor and reviewers for their valuable comments and suggestions to improve the paper.

Conflict of interest

QL was employed by Shandong Xinhua'an Information Technology Co., Ltd.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Ahmed, I., Yadav, P. K. (2023). A systematic analysis of machine learning and deep learning based approaches for identifying and diagnosing plant diseases. Sustain. Operations Comput. 4, 96–104. doi: 10.1016/j.susoc.2023.03.001

Crossref Full Text | Google Scholar

Anim-Ayeko, A. O., Schillaci, C., Lipani, A. (2023). Automatic blight disease detection in potato (solanum tuberosum l.) and tomato (solanum lycopersicum, l. 1753) plants using deep learning. Smart Agric. Technol. 4, 100178. doi: 10.1016/j.atech.2023.100178

Crossref Full Text | Google Scholar

Chang, C.-Y., Lai, C.-C. (2024). Potato leaf disease detection based on a lightweight deep learning model. Mach. Learn. Knowledge Extraction 6, 2321–2335. doi: 10.3390/make6040114

Crossref Full Text | Google Scholar

Chen, C.-J., Huang, Y.-Y., Li, Y.-S., Chen, Y.-C., Chang, C.-Y., Huang, Y.-M. (2021). Identification of fruit tree pests with deep learning on embedded drone to achieve accurate pesticide spraying. IEEE Access 9, 21986–21997. doi: 10.1109/ACCESS.2021.3056082

Crossref Full Text | Google Scholar

Chen, J., Kao, S.-h., He, H., Zhuo, W., Wen, S., Lee, C.-H., et al. (2023). “Run, don’t walk: chasing higher flops for faster neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) (Vancouver, BC, Canada: IEEE), 12021–12031. doi: 10.1109/CVPR52729.2023.01157

Crossref Full Text | Google Scholar

Faisal, H. M., Aqib, M., Mahmood, K., Safran, M., Alfarhood, S., Ashraf, I. (2024). A customized convolutional neural network-based approach for weeds identification in cotton crops. Front. Plant Sci. 15. doi: 10.3389/fpls.2024.1435301

PubMed Abstract | Crossref Full Text | Google Scholar

Gaikwad, V. P., Musande, V. (2023). Advanced prediction of crop diseases using cetalatran-optimized deep knn in multispectral imaging. Traitement Du Signal 40 1093. doi: 10.18280/ts.400325

Crossref Full Text | Google Scholar

Gao, L., Zhao, X., Yue, X., Yue, Y., Wang, X., Wu, H., et al. (2024). A lightweight yolov8 model for apple leaf disease detection. Appl. Sci. 14, 6710. doi: 10.3390/app14156710

Crossref Full Text | Google Scholar

Harakannanavar, S. S., Rudagi, J. M., Puranikmath, V. I., Siddiqua, A., Pramodhini, R. (2022). Plant leaf disease detection using computer vision and machine learning algorithms. Global Transitions Proc. 3, 305–310. doi: 10.1016/j.gltp.2022.03.016

Crossref Full Text | Google Scholar

Hou, J., Yang, C., Hou, B. (2023). Detecting diseases in apple tree leaves using fpn–isresnet–faster rcnn. Eur. J. Remote Sens. 56, 2186955. doi: 10.1080/22797254.2023.2186955

Crossref Full Text | Google Scholar

Islam, M. M., Adil, M. A. A., Talukder, M. A., Ahamed, M. K. U., Uddin, M. A., Hasan, M. K., et al. (2023). Deepcrop: Deep learning-based crop disease prediction with web application. J. Agric. Food Res. 14, 100764. doi: 10.1016/j.jafr.2023.100764

Crossref Full Text | Google Scholar

Khan, A. I., Quadri, S., Banday, S., Shah, J. L. (2022). Deep diagnosis: A real-time apple leaf disease detection system based on deep learning. Comput. Electron. Agric. 198, 107093. doi: 10.1016/j.compag.2022.107093

Crossref Full Text | Google Scholar

Li, T., Zhang, L., Lin, J. (2024). Precision agriculture with yolo-leaf: advanced methods for detecting apple leaf diseases. Front. Plant Sci. 15. doi: 10.3389/fpls.2024.1452502

PubMed Abstract | Crossref Full Text | Google Scholar

Li, Y., Wang, H., Dang, L. M., Sadeghi-Niaraki, A., Moon, H. (2020). Crop pest recognition in natural scenes using convolutional neural networks. Comput. Electron. Agric. 169, 105174. doi: 10.1016/j.compag.2019.105174

Crossref Full Text | Google Scholar

Liu, B., Huang, X., Sun, L., Wei, X., Ji, Z., Zhang, H. (2024a). Mcdcnet: Multi-scale constrained deformable convolution network for apple leaf disease detection. Comput. Electron. Agric. 222, 109028. doi: 10.1016/j.compag.2024.109028

Crossref Full Text | Google Scholar

Liu, D., Xu, J., Li, X., Zhang, F. (2024b). Green production of apples delivers environmental and economic benefits in China. Plant Commun. 5, 101006. doi: 10.1016/j.xplc.2024.101006

PubMed Abstract | Crossref Full Text | Google Scholar

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., et al. (2016). “Ssd: Single shot multibox detector,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I, vol. 14. (Amsterdam, Netherlands: Springer), 21–37. doi: 10.1007/978-3-319-46448-0_2

Crossref Full Text | Google Scholar

Lv, M., Su, W.-H. (2024). Yolov5-cbam-c3tr: an optimized model based on transformer module and attention mechanism for apple leaf disease detection. Front. Plant Sci. 14. doi: 10.3389/fpls.2023.1323301

PubMed Abstract | Crossref Full Text | Google Scholar

Ma, B., Hua, Z., Wen, Y., Deng, H., Zhao, Y., Pu, L., et al. (2024). Using an improved lightweight yolov8 model for real-time detection of multi-stage apple fruit in complex orchard environments. Artif. Intell. Agric. 11, 70–82. doi: 10.1016/j.aiia.2024.02.001

Crossref Full Text | Google Scholar

Menghani, G. (2023). Efficient deep learning: A survey on making deep learning models smaller, faster, and better. ACM Computing Surveys 55, 1–37. doi: 10.1145/3578938

Crossref Full Text | Google Scholar

Ouyang, D., He, S., Zhang, G., Luo, M., Guo, H., Zhan, J., et al. (2023). “Efficient multi-scale attention module with cross-spatial learning,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (Rhodes Island, Greece: IEEE), 1–5. doi: 10.1109/ICASSP49357.2023.10096516

Crossref Full Text | Google Scholar

Peng, H., Yu, S. (2021). A systematic iou-related method: Beyond simplified regression for better localization. IEEE Trans. Image Process. 30, 5032–5044. doi: 10.1109/TIP.2021.3077144

PubMed Abstract | Crossref Full Text | Google Scholar

Ramesh, S., Hebbar, R., Niveditha, M., Pooja, R., Shashank, N., Vinod, P., et al. (2018). “Plant disease detection using machine learning,” in 2018 International conference on design innovations for 3Cs compute communicate control (ICDI3C) (Bangalore, India: IEEE), 41–45. doi: 10.1109/ICDI3C.2018.00017

Crossref Full Text | Google Scholar

Reddy, S. R., Varma, G. S., Davuluri, R. L. (2023). Resnet-based modified red deer optimization with dlcnn classifier for plant disease identification and classification. Comput. Electrical Eng. 105, 108492. doi: 10.1016/j.compeleceng.2022.108492

Crossref Full Text | Google Scholar

Redmon, J. (2016). “You only look once: Unified, real-time object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR). (Las Vegas, NV, USA: IEEE), 779–788. doi: 10.1109/CVPR.2016.91

Crossref Full Text | Google Scholar

Ren, S., He, K., Girshick, R., Sun, J. (2016). Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39, 1137–1149. doi: 10.1109/TPAMI.2016.2577031

PubMed Abstract | Crossref Full Text | Google Scholar

Shao, Y., Yang, W., Wang, J., Lu, Z., Zhang, M., Chen, D. (2024). Cotton disease recognition method in natural environment based on convolutional neural network. Agriculture 14, 1577. doi: 10.3390/agriculture14091577

Crossref Full Text | Google Scholar

Soujanya, K., Jabez, J. (2021). “Recognition of plant diseases by leaf image classification based on improved alexnet,” in 2021 2nd International Conference on Smart Electronics and Communication (ICOSEC) (Trichy, India: IEEE), 1306–1313. doi: 10.1109/ICOSEC51865.2021.9591809

Crossref Full Text | Google Scholar

Sun, H., Xu, H., Liu, B., He, D., He, J., Zhang, H., et al. (2021). Mean-ssd: A novel real-time detector for apple leaf diseases using improved light-weight convolutional neural networks. Comput. Electron. Agric. 189, 106379. doi: 10.1016/j.compag.2021.106379

Crossref Full Text | Google Scholar

Sun, J., Zhang, J., Gao, X., Wang, M., Ou, D., Wu, X., et al. (2022). Fusing spatial attention with spectral-channel attention mechanism for hyperspectral image classification via encoder–decoder networks. Remote Sens. 14, 1968. doi: 10.3390/rs14091968

Crossref Full Text | Google Scholar

Tian, D., Han, Y., Wang, S., Chen, X., Guan, T. (2022). Absolute size iou loss for the bounding box regression of the object detection. Neurocomputing 500, 1029–1040. doi: 10.1016/j.neucom.2022.06.018

Crossref Full Text | Google Scholar

Wan, D., Lu, R., Hu, B., Yin, J., Shen, S., Lang, X., et al. (2024). Yolo-mif: Improved yolov8 with multi-information fusion for object detection in gray-scale images. Advanced Eng. Inf. 62, 102709. doi: 10.1016/j.aei.2024.102709

Crossref Full Text | Google Scholar

Wang, C.-Y., Bochkovskiy, A., Liao, H.-Y. M. (2023). “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (Vancouver, Canada: IEEE), 7464–7475. doi: 10.48550/arXiv.2207.02696

Crossref Full Text | Google Scholar

Wang, J., Xu, C., Yang, W., Yu, L. (2021). A normalized gaussian wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389. doi: 10.48550/arXiv.2110.13389

Crossref Full Text | Google Scholar

Wang, S., Li, Q., Yang, T., Li, Z., Bai, D., Tang, C., et al. (2024a). Lsd-yolo: Enhanced yolov8n algorithm for efficient detection of lemon surface diseases. Plants 13, 2069. doi: 10.3390/plants13152069

PubMed Abstract | Crossref Full Text | Google Scholar

Wang, Y., Wang, W., Li, Y., Jia, Y., Xu, Y., Ling, Y., et al. (2024b). An attention mechanism module with spatial perception and channel information interaction. Complex Intelligent Syst. 10, 5427–5444. doi: 10.1007/s40747-024-01445-9

Crossref Full Text | Google Scholar

Wani, N. A., Mishra, U. (2022). An integrated circular economic model with controllable carbon emission and deterioration from an apple orchard. J. Cleaner Production 374, 133962. doi: 10.1016/j.jclepro.2022.133962

Crossref Full Text | Google Scholar

Wei, D., Chen, J., Luo, T., Long, T., Wang, H. (2022). Classification of crop pests based on multi-scale feature fusion. Comput. Electron. Agric. 194, 106736. doi: 10.1016/j.compag.2022.106736

Crossref Full Text | Google Scholar

Yan, C., Liang, Z., Yin, L., Wei, S., Tian, Q., Li, Y., et al. (2024). Afm-yolov8s: An accurate, fast, and highly robust model for detection of sporangia of plasmopara viticola with various morphological variants. Plant Phenomics 6, 246. doi: 10.34133/plantphenomics.0246

PubMed Abstract | Crossref Full Text | Google Scholar

Zeng, W., Li, H., Hu, G., Liang, D. (2022). Lightweight dense-scale network (ldsnet) for corn leaf disease identification. Comput. Electron. Agric. 197, 106943. doi: 10.1016/j.compag.2022.106943

Crossref Full Text | Google Scholar

Zhang, J., Chen, Z., Yan, G., Wang, Y., Hu, B. (2023a). Faster and lightweight: An improved yolov5 object detector for remote sensing images. Remote Sens. 15, 4974. doi: 10.3390/rs15204974

Crossref Full Text | Google Scholar

Zhang, Y.-F., Ren, W., Zhang, Z., Jia, Z., Wang, L., Tan, T. (2022). Focal and efficient iou loss for accurate bounding box regression. Neurocomputing 506, 146–157. doi: 10.1016/j.neucom.2022.07.042

Crossref Full Text | Google Scholar

Zhang, Y., Zhou, G., Chen, A., He, M., Li, J., Hu, Y. (2023b). A precise apple leaf diseases detection using bctnet under unconstrained environments. Comput. Electron. Agric. 212, 108132. doi: 10.1016/j.compag.2023.108132

Crossref Full Text | Google Scholar

Zheng, Q., Chen, Z., Liu, H., Lu, Y., Li, J., Liu, T. (2023). Msranet: Learning discriminative embeddings for speaker verification via channel and spatial attention mechanism in alterable scenarios. Expert Syst. Appl. 217, 119511. doi: 10.1016/j.eswa.2023.119511

Crossref Full Text | Google Scholar

Zhong, H., Lv, Y., Yuan, R., Yang, D. (2022). Bearing fault diagnosis using transfer learning and self-attention ensemble lightweight convolutional neural network. Neurocomputing 501, 765–777. doi: 10.1016/j.neucom.2022.06.066

Crossref Full Text | Google Scholar

Zhu, J., Wu, Y., Ma, T. (2024). Multi-object detection for daily road maintenance inspection with uav based on improved yolov8. IEEE Trans. Intelligent Transportation Syst. 25, 16548–16560. doi: 10.1109/tits.2024.3437770

Crossref Full Text | Google Scholar