The apple leaf disease data utilized in the paper have been selected from the Phytopathology 2021 (FGVC8) dataset and the AppleLeaf9 dataset. The images in the datasets are all derived from outdoor scenes. Seven common diseases have been chosen for the study. Frog_eye_leaf_spot, Powdery_mildew, Rust, and Scab were selected from FGVC8. Alternaria leaf spot, Grey spot, and Mosaic were selected from AppleLeaf9. Moreover, under natural conditions Frog_eye_leaf_spots are mixed with Rust and Scab to form two categories of disease occurrence scenarios, respectively. In total, there are seven disease names and nine disease categories. The specific number of images in each category is shown in Table 1 . The representative images of different disease categories are shown in Figure 1 .

All collected images were labeled using the LabelImg tool and saved in an XML file. In addition, the images in the datasets are enhanced with changes such as rotation, brightness, contrast, and addition of noise. Finally, there are 18320 images in the datasets, and the specific number of images in each category is shown in Table 1 . During the datasets labeling process, this study has assumed all the browning areas on the leaves to be spots caused by pests and diseases. However, this study recognizes that browning may also be caused by other reasons such as environmental factors or human intervention, and this simplifying assumption may limit the comprehensiveness of the labeling and the generalization ability of the model. In order to improve the accuracy and rigor of the study, this study plans to introduce a detailed comparative analysis of types of browning not caused by pests and diseases in future studies. This will help understand the multiple causes of browning and make appropriate adjustments in datasets labeling to ensure that the model can accurately identify and distinguish among various browning types and improve overall detection. The links to the datasets used in this study are provided below: https://drive.google.com/drive/folders/1MRfK5eOm5-6KZTngPzpzjp9gx1NyEvZY?usp=sharing.

As illustrated in Figure 1 , apple leaf disease detection faces several challenges. Firstly, different categories of leaf diseases have different shapes and sizes, which makes feature extraction more difficult. Secondly, most leaf diseases are small and densely distributed, which increases the difficulty of the localization process. Finally, under outdoor conditions, natural light and raindrops may interfere with leaf disease recognition.

A series of different scenes were selected as part of the experimental datasets for this study. The dark scene simulated the detection of leaf diseases in a low light or night environment. The rainy scene symbolized the effects of rain on the leaf surface, such as raindrop shading and water droplet retention. The lighting scene emphasizes the situation of leaves under direct sunlight or bright light. The dense scene contains a large number of diseases, which is closer to the leaf disease situation in real farms and provides a more rigorous testing environment for evaluating the performance of the detection algorithms. The multiple leaves scene and the two-spots scene further increase the complexity of the scenarios by taking into account multiple leaves and the interactions between two spots on the leaves. By conducting experiments in these different scenarios, the applicability of the proposed leaf disease detection method in practical applications can be comprehensively evaluated and more reliable technical support can be provided for agricultural production.

With the aim of achieving a lightweight model on apple leaf disease detection to make it more convenient to be applied to embedded devices, the paper proposes a lightweight method and names it LCGSC-YOLO. Figures 2A, B show the detailed framework of YOLOv5 and the proposed LCGSC-YOLO in the paper, respectively.

YOLO is chosen as the main framework in this study mainly because of its efficiency and real-time performance. YOLO is able to predict both bounding box and class probabilities of targets in a single network, which allows it to perform well when dealing with complex backgrounds and dense targets while maintaining low computational requirements and fast inference speed. Some other target detection framework, although advantageous in different aspects, such as SSD performs better in terms of speed but may not be as good as YOLO when dealing with small targets and complex backgrounds. Faster R-CNN, although it performs well in terms of detection accuracy, its two-phase structure results in slower inference and higher computational requirements, which limits its application in real-time detection. Taking these factors into consideration, YOLO is chosen in this study to meet our requirements for real-time performance and computational efficiency, while effectively handling complex backgrounds and high-density targets. The YOLOv5 version is chosen because it has demonstrated excellent stability and maturity in the field of target detection, providing a solid foundation for innovation in this study.

YOLOv5 is made up of four parts: the input layer, the backbone network, the neck network and the prediction head. As presented in Figure 2A , it can be seen that the backbone network of YOLOv5 stacks numerous Conv and C3 modules. The Spatial Pyramid Pooling-Fast (SPPF) module is utilized to capture multi-scale target information and then connects to the neck network (Zhang et al., 2023). In the neck network, besides the Conv and C3 modules, the Concat module is employed to aggregate the feature maps of different layers, thus reducing feature map information loss. The detection head module mainly performs multi-scale target detection of feature maps (Sun et al., 2022).

Due to the large number of Conv and C3 modules in the original YOLOv5 framework, it is difficult to embed it for utilizing in mobile devices (Xu and Wang, 2023). Therefore, an efficient LCGSC-YOLO for apple leaf disease detection is proposed in the work. The framework of LCGSC-YOLO is shown as Figure 2B . Compared to YOLOv5, the main innovations of LCGSC-YOLO are described below: 1) LCNet is used to reconstruct the backbone network and is categorized into LCNet-3 and LCNet-5 according to the convolutional kernel size. LCNet greatly reduces the number of parameters and computations of the model. 2) In the neck network, we are utilizing the GSConv module and the VOVGSCSP module to replace the Conv and C3 modules from the YOLO framework. 3) The coordinate attention(CA) is inserted at the tail of the backbone and after each VOVGSCSP module, which alleviates the problem of detection accuracy reduction caused by the lightweighting of the model.

In the next section, the effectiveness of each module will be analyzed step by step. Firstly, this study demonstrates that the lightweight design of LCNet effectively reduces the computational complexity through theory and formulas. Similarly, this study analyzes that GSConv optimizes the convolution operation by mixing convolution kernels, which further reduces the computation. Secondly, the introduction of CA attention mechanism makes the accuracy of the model improved. Finally, the experimental results show that the combination of LCNet and GSConv makes a significant reduction in the number of parameters and the number of Floating Point Operations (FLOPs) of the model, while the CA attention mechanism further enhances the performance and accuracy of the model.

The experiments are performed in Ubuntu system with an Intel Xeon(R) Silver 4214R CPU@2.40 GHz x48 processor, 128 GB of RAM, a graphics card NVIDIA Corporation TU102GL [Quadro RTX 8000], CUDA 12.2, Pytorch 2.1.0, Python 3.9.18. The hyperparameters of the experiments are set as follows: the Epochs of the model are set to 300, the Initial learning rate of the model is given as 0.01, the model optimizer is selected as SGD, and the Batch size is fixed to 32. All apple leaf disease data are classified by a ratio of 7 to 2 to 1 as training set, validation set, and test set.

Aiming to objectively evaluate the validity of the experimental results, the paper chooses mean accuracy precision (mAP), precision (P), recall (R) as the objective evaluation metrics of the experiment (Hossin and Sulaiman, 2015). The mAP denotes the sum of the mean accuracies of all categories divided by all categories. The P means the ratio of the actual number of positive samples in the predicted sample to the number of all positive samples. The R represents the ratio of the number of actual positive samples in the predicted sample to all predicted samples. The assessment metrics were calculated according to the following Equations 9–12.

TP (True Positive) indicate the number of positive samples that the model correctly predicts as positive. FP (False Positive) expresses the number of negative samples that the model incorrectly predicts as positive. FN (False Negative) signifies the number of positive samples that the model incorrectly predicts as negative (Zhu et al., 2023).

Moreover, the number of parameters (Para) and FLOPs are employed to assess the complexity of the model (Justus et al., 2018). Assume that the convolution kernel size is A k × A k , A i × A i × C 1 is the size of the input feature map, and the size of the output feature map is A i × A i × C 2 . The fewer the number of parameters and FLOPs, and the lower the complexity of the model, which means it is more suitable to be applied in resource-constrained embedded devices.

The number of depth-wise separable convolution parameters is indicated as Equation 13:

The calculation formula of FLOPs for depth-wise separable convolution is shown in Equation 4.

Finally, frames per second (FPS) is utilized to evaluate the inference speed of the model (Kiani Galoogahi et al., 2017), larger values of FPS indicate that more data is processed within the same time.

A series of testing experiments validate the effectiveness of the different lightweighting modules proposed in the paper. In order to ensure the generalization ability of the model, this study has adopted the method of dividing the validation set. The datasets have been divided according to 70% for training, 20% for validation, and 10% for testing. With this validation strategy, this study evaluated the performance of the model on unseen data. Test 1 indicates YOLOv5s, which is treated as a baseline model. Test 2 means reconstructing the backbone of the YOLO by LCNet. Test 3 demonstrates the adoption of the GSConv module and the VOVGSCSP module to replace the Conv and C3 modules in the YOLO framework, respectively. Test 4 indicates a combination of improvements from Test 2 and Test 3. The results of the objective assessment of the model detection performance with different improvement modules are listed in Table 2 .

The conclusion can be obtained by analyzing the experimental results of Test 1 and Test 2. In comparison for YOLOv5s, the number of parameters and FLOPs of YOLO-LCNet are decreased by 48.22% and 60.63%, respectively, and increases the FPS by 60%. It demonstrates that utilizing the LCNet module to reconstruct the backbone network of the YOLO framework can dramatically decrease the complexity of the model and can quite obviously increase the inference speed of the model. From Test 1 and Test 3, the modified neck network of YOLO with GSConv_VOVGSCSP module (YOLO+GS_VOV) reduces the number of parameters and FLOPs by 0.99M and 1.8G, respectively, while mAP is reduced by only 0.5%. The consequences confirm that the introduction of the proposed GSConv_VOVGSCSP (GS_VOV) module into the neck network not only makes the model more lightweight but also has little effect on the feature fusion capability among the different network layers in the neck network. Based on the experimental results of Test 4, it can be observed that the number of model parameters and FLOPs in the YOLO framework introduced by LCNet together with the GS_VOV module are only 37.12% and 35% of those in YOLOv5s, respectively. The model inference speed attained 60 FPS, which is 1.71 times faster than that of YOLOv5s. The above discussion and analysis demonstrate the effectiveness of the proposed method.

It is clear from Table 2 that the lightweighting of the model inevitably brings about the problem of detection accuracy degradation. Therefore, it is very necessary to improve the detection accuracy as much as possible without bringing in higher computations.

We take YOLO+LCNet+GS_VOV as the baseline network and verify the influence in model performance after introducing different attention modules. Test 4 is YOLO+LCNet+GS_VOV, which combines the LCNet, GSConv and VOVGSCSP modules. Test 5 represents the SE module that is added to Test 4, Test 6 indicates the CBAM module that is introduced to Test 4, and Test 7 is the proposed LCGSC-YOLO, which integrates the CA module from Test 4. The objective evaluation results of introducing different attention modules on the performance of YOLO+LCNet+GS_VOV are reflected in Table 3 , respectively.

Compared to Test 4, the results of Test 5, Test 6, and Test 7 demonstrate that adding the attention mechanism to YOLO+LCNet+GS_VOV can enhance the detection performance of the model. Although all the listed attention mechanisms improve detection accuracy, they simultaneously increasing the number of parameters and decreasing the model inference speed. However, the task at this stage is to maximize accuracy within a limited range of parameters variation. The key is to reach a balance between lightweighting and accuracy. Therefore, in this paper, CA is chosen at this stage because of the desire to increase the detection accuracy as much as possible. As can be observed in Test 4 and Test 7, the mAP of LCGSC-YOLO is 1.3% higher than that of YOLO+LCNet+GS_VOV, while the number of parameters and computations only increase by 0.35 M and 1.1 G, respectively. Even though there is some computational cost associated with the approach, it has very little effect on the inference speed of the model. LCGSC-YOLO inference speed is reduced by only 7 FPS. With the above results, the CA attention mechanism can effectively alleviate the problem of detection accuracy degradation caused by the model lightweighting. Therefore, the addition of CA attention mechanism to the proposed lightweight model can make the model performance more excellent.

After introducing the CA attention module in the lightweight model, the research further conducts ablation experiments by combining CA with above modules of lightweight to verify the independence of the CA module and to prove that there is no dependency among the modules. The results in Table 4 show that Test 8 represents YOLO+GS_VOV+CA, Test 9 denotes YOLO+LCNet+CA, and Test 10 indicates YOLOv5s+CA. Compared with Test 3 in Table 2 , Test 8 shows the increase in accuracy and recall after the introduction of the CA module, proving that the CA module plays an active role in model performance optimization. Similarly, the comparison between Test 9 and Test 2 shows that the introduction of the CA module improves the detection performance, which mAP improves to 95.9%. Test 10 further demonstrates the effectiveness of the CA module in the YOLOv5s model, with mAP reaching 97.1%. The above experimental results validate that the CA module is able to produce improvements on multiple lightweight modules. However, although the detection performance is improved, this improvement is accompanied by both increase in the number of parameters and computational complexity. Therefore, there is a balance between accuracy and model complexity.

As a conclusion, LCGSC-YOLO has several advantages. Firstly, the model with small number of parameters do not require a large amount of storage space. Secondly, the model is low computation and can be run with limited hardware resources. Finally, the model training and inference speed is fast and can process the data quickly. Therefore, LCGSC-YOLO is more suitable to be deployed in embedded devices for detecting apple leaf diseases.

In this subsection, aiming to validating the performance of different backbone networks, the backbone of the YOLO framework is reconstructed by the current mainstream lightweight modules. YOLO-MN3 means the backbone of YOLO is constructed by employing the MobileNetv3 module (Howard et al., 2019), and YOLO-SN2 illustrates that it is reconstructed by applying the basic modules of ShuffleNetv2 (Ma et al., 2018), YOLO-GN and YOLO-EN2 denote that the YOLO backbone is composed of modules applying GhostNet (Khan et al., 2022) and EfficientNetv2 (Tian et al., 2020), respectively, and the backbone with the LCNet is known as YOLO-LCNet. Table 5 lists the results of the test experiments for different lightweight backbone networks.

As can be noticed from Table 5 , all of listed different lightweight modules can to some degree decrease the number of parameters and computations of the model, but inevitably causes a loss of accuracy. Compared with YOLO-MN3, YOLO-SN2, YOLO-GN and YOLO-EN2, YOLO-LCNet has lower model complexity and quicker inference speed. Specifically, the mAP of YOLOv5 is only 1.3% higher compared to YOLO-LCNet, but the number of parameters and FLOPs of YOLO-LCNet are 48.22% and 60.63% lower than that of YOLOv5, respectively. In addition, compared to YOLOv5, the model inference speed of YOLO-LCNet is 60% faster than it. In a word, Using LCNet to reconstruct the backbone network of YOLO is a better choice to reduce model complexity and enhance model inference speed.

In this section, the experimental results of the proposed method are compared with other methods related to leaf disease detection. Specifically, INAR-SSD (Jiang et al., 2019) as a detection model for apple leaf disease detection with the ALDD dataset. BTC-YOLOv5s (Li et al., 2023) and MGA-YOLO (Wang et al., 2022) are improved lightweight apple leaf disease detection models based on the FGVC8 datasets. Khan et al. employ YOLOv4 to apple leaf disease detection (Khan et al., 2022). The experimental results of different models are presented in Table 6 . The reason for selecting the above comparison methods is that they have some similarities with the method proposed in this study in terms of theoretical and technical characteristics. For example, both INAR-SSD and YOLOv4 belong to the classical methods in the field of target detection.BTC-YOLOv5s and MGA-YOLO are improved versions based on a lightweight target detection model, and their design ideas and technical features have some similarities with the method proposed in this study. In addition, these comparative methods use similar datasets when dealing with leaf disease detection tasks, and all attempt to address common challenges in leaf disease detection, such as light variations and leaf disease morphological diversity. Therefore, by comparing with these methods, the innovations and usefulness of the method proposed in this study can be better assessed in the context of current research hot spots and technological trends, as well as its strengths and limitations in solving the leaf disease detection problem.

As shown in Table 6 , the proposed LCGSC-YOLO model has fewer number of parameters and computations compared with other experimental models. Moreover, it also shows excellent performance in terms of detection accuracy and inference speed. In comparison with INAR-SSD, LCGSC-YOLO achieves a 10.4% increase in detection accuracy and a 46 FPS improvement in inference speed. In addition, the proposed method has only 12.53% of the number of parameters and 7.47% of the FLOPs of INAR-SSD, respectively. Meanwhile, the comparison results with BTC-YOLOv5s and MGA-YOLO show that the detection accuracies of BTC-YOLOv5s and MGA-YOLO are 5.4% and 1.7% less than that of LCGSC-YOLO, respectively. Besides, the model inference speed of LCGSC-YOLO are 4.41 and 2.52 times faster than that of them, respectively. However, the parameter amount and FLOPs of LCGSC-YOLO are only 18.73% and 12.60% of those of BTC-YOLOv5s, and about a quarter of those of MGA-YOLO. In comparison of YOLOv4, LCGSC-YOLO has 57.85M and 38.1G fewer parameters and FLOPs, respectively, while the model inference speed of LCGSC-YOLO is 39 FPS more than YOLOv4. The above analysis results illustrate that the proposed method is superior to the comparative experimental methods in terms of comprehensive performance.

Figure 6 shows the radar charts of the different model experimental results. it can be concluded that INAR-SSD has the largest FLOPs, YOLOv4 has the largest number of parameters, and the mean average precision, precisions, and recalls of the different modelling methods are almost overlapping. In addition, The FPS of LCGSC-YOLO is obviously superior to the other methods, while the number of parameters and FLOPs are lower than those of the comparative experimental methods. In terms of the area surrounded by the test results of different models, LCGSC-YOLO mainly occupies the left area of the figure, which implies that LCGSC-YOLO has better detection performance while having fewer number of parameters and FLOPs.

As illustrated in Figure 7 , a 3D bar chart is drawn based on the test results of different models, which can visualize the comparison results of different models. It comes to a conclusion that the test results of different models do not show much difference in terms of mean average precision, precision, and recall. However, it can be observed that from INAR-SSD to LCGSC-YOLO are gradually increasing and decreasing in terms of the performance of FPS and FLOPs, respectively. It proclaims that the computational amount of the LCGSC-YOLO is gradually decreasing, and the inference speed is continuously increasing. Overall, compared with other models, LCGSC-YOLO has the lowest number of parameters and computational amount as well as the fastest inference speed.

To visualize the detection performance of the LCGSC-YOLO, Figure 8 shows the detection results of INAR-SSD, BTC-YOLOv5s, YOLOv4 and MGA-YOLO and LCGSC-YOLO on different disease images, respectively. To improve readability, the detection information in Figure 8 substitutes specific letters for disease names. A represents Scab, B represents Rust, C denotes Powdery_mildew, D denotes Frog_eye_leaf_spot, E indicates Alternaria leaf spot, and F indicates Grey spot, G means Mosaic. As shown in Figure 8 , from the overall detection results, the proposed LCGSC-YOLO outperforms other algorithms in detecting different kinds of diseases. In particular, LCGSC-YOLO has better detection performance in detecting small diseases. To be specific, as indicated by the red circle from the Scab detection result images, it can be seen that LCGSCYOLO can separately detect adjacent disease regions, while other algorithms recognize them as a single disease. Moreover, it is apparent from the result images of Rust diseases that LCGSCYOLO can detect small disease areas at the edge positions. In addition, for the three diseases Frog_eye_leaf_spot, Alternaria leaf spot, and Grey spot, all methods have varying degrees of miss detection. In these three disease categories, it was difficult for all methods to identify all diseases due to the simultaneous presence of diseases in multiple leaves in the same scene. For Powdery_mildew and Mosaic, the detection results of all methods were almost the same.

To further illustrate the superiority of LCGSC-YOLO, Figure 9 compares the model detection capabilities of INAR-SSD, BTC-YOLOv5s, YOLOv4, MGA-YOLO, and LCGSC-YOLO on specific scenarios. The scenes from top to bottom are dark, rainy, strong lighting, multiple leaves, two spots and dense scenes. As displayed in Figure 9 , for disease images in dark scenes, the other four algorithms did not recognize small diseases, while LCGSC-YOLO was able to recognize them. For images of rainy scenes, LCGSC-YOLO accurately detected Grey spot that were difficult to recognize due to rainfall reflection, but the other four algorithms identified the disease spot as the same as adjacent disease spots. In the detection results of two spots scenes and multiple leaves scenes, there were varying degrees of missed detections. Due to the small size and dispersion of all diseases, it is difficult for all methods to identify all diseases. In lighting scenes, INAR-SSD incorrectly identifies a disease in the light. In addition, in dense scenes, LCGSC-YOLO can identify adjacent diseases separately. From the aforementioned analysis, it can be concluded that LCGSC-YOLO also has comparably equally excellent detection performance in special scenarios.

In general, by combining the detection capabilities of different models in different scenarios in Figures 8 , 9 , it is obvious that LCGSC-YOLO shows excellent performance in detecting apple leaf diseases. The model utilizes fewer parameters and lower computational effort to achieve efficient leaf disease detection, and especially excels in model lightweighting. However, despite the obvious advantages of LCGSC-YOLO in terms of lightweighting, it is still important to note that its performance may be degraded when dealing with complex scenarios such as dense diseases, multiple leaves or two spots. In addition, the model showed missed detection in terms of detecting tiny diseases, which requires further optimization of the method. Therefore, in the future, the study will continue to explore the optimization of the LCGSC-YOLO model to improve its robustness and adaptability, and to further optimize its detection accuracy for more comprehensive and stable apple leaf disease detection.