In this study, we utilized the highly regarded PlantVillage dataset (Davi et al., 2015) (Hughes and Marcel, 2015), which includes 38 categories of plant leaf diseases and a total of 54,305 images. This dataset consists of 36,219 images in the training set, 9,036 images in the validation set, and 9,050 images in the test set, with example images shown in Figure 1.A. This dataset was chosen for its widespread use as a benchmark in plant disease recognition, providing a robust foundation for initial model validation and comparison with existing literature. Each image in this dataset has been rigorously validated and annotated by experienced plant pathologists, ensuring a high level of accuracy and reliability.

To further validate the advancements and generalizability of our method, we also utilized the AI Challenger 2018 Dataset (Wu et al., 2019) (Wu et al., [[NoYear]]), which comprises 61 categories of plant leaf diseases at varying degrees of severity, totaling 36,075 images. This dataset includes 25,252 images in the training set, 6,289 images in the validation set, and 4,534 images in the test set, with example images displayed in Figure 1.B. Its highly imbalanced category labels make recognition inherently more challenging. This dataset serves as a critical test for model robustness under difficult conditions. In this study, we treated it as an ablation dataset and did not apply any augmentation techniques to specifically assess the model’s inherent performance on imbalanced data without external data manipulation.

We propose a novel Lightweight Multi-Path Pruning Method (LMP-PM), as illustrated in Figure 2. First, we design an original model, assuming the input feature map is represented by X∈RC×A×B, with C indicating the number of input channels, and A and B representing the height and width of the input feature map, respectively. We define W as the weight matrix of the original convolutional layer; therefore, the output of the original model can be expressed by formula (1):

where f represents the index of the output channels, and Xi,a+j,b+k is the input feature map. K denote the size of the convolutional kernel, with its center position located at ( K2, K2). Considering the symmetry of the convolutional kernel, indices from −K2 to K2 can cover the entire kernel. Next, we proceed to lightweight the original model by reducing the number of input feature map channels to CN. N is the pruning parameter, and its size is determined based on the complexity of the original model. The specific steps for pruning are as follows: First, we iterate through all layers, sequentially extracting each standard convolutional layer from the original model. For the current convolutional layer, we retrieve its output channel count, denoted as C. Next, a custom pruning parameter, N, is introduced. The value of this parameter N is typically determined based on the original model’s complexity and the desired degree of lightweighting. Subsequently, we check if the current layer’s output channel count satisfies the pruning condition, specifically C/N > 1. If this condition is not met, the layer is not pruned, thereby preventing excessive compression that could lead to a drastic performance drop. If the pruning condition is met, the layer’s output channel count C is updated to C = C/N. This action, in turn, reduces the computational burden on this layer and subsequent layers. Finally, these steps are repeated until all convolutional layers have been traversed and their channel counts adjusted according to the pruning conditions. We define the new convolutional weights as W1, therefore, the output feature map can be expressed as formula (2):

where the new convolutional weights W1∈RF1×CN×K×K, f1 serve as indices for the new output channels. Following this, we perform the next lightweight operation by replacing the standard convolution with depthwise (DW) convolution, as outlined in (Howard et al., 2017). Subsequently, the process entails re-traversing all layers within the channel-pruned model to determine whether the current layer is a standard convolutional layer. If so, its critical parameters—namely, in_channels, out_channels, kernel_size, stride, and padding—are extracted. These extracted parameters are then utilized to construct a depthwise convolutional layer. At this point, the input feature map, denoted as Y1, processed to obtain formula (3):

where the weights of the depthwise convolution are Wd∈RCN×1×K×K, where j and k serve as the indices for the convolution operation. Immediately following this, we apply pointwise convolution to Y2, resulting in formula (4):

Then, to balance the performance loss associated with the two lightweight operations, we convert the single-path model into a multi-path model by controlling the path expansion factor E. The specific steps are as follows: First, we examine the value of the custom path expansion factor, E. If E ≤ 1, multi-path expansion is not performed; instead, the lightweight single-path model is utilized directly. However, if E > 1, an empty list is initialized to store the multiple model branches that will be subsequently created. A loop is then initiated, with the loop variable ‘i’ iterating from 1 to E. In each iteration, a deep copy of the structurally-transformed lightweight single-path model is performed. This implies copying not only the model’s architecture but also all of its current weights. Afterwards, the weights of all branches except the original one are randomly initialized. The duplicated model instance is then appended to the pre-initialized list of branches. Upon the loop’s completion, the number of paths is determined based on the specific task, resulting in formula (5):

Finally, we concatenate all the paths, after lightweight optimization, the final output results of the model are shown in formula (6):

We propose an original model (OMNet), whose structure is illustrated in Figure 3A. OMNet was designed to deliver high performance and high complexity, serving as the foundation for subsequent lightweight processing in LMP-PM. OMNet comprises fundamental components such as convolutional layers, pooling layers, activation functions, and normalization functions, while also integrating residual modules (He et al., 2016), SE modules (Hu et al., 2018), and TBP blocks. The overall design approach of OMNet involves first downsampling the input image twice, reducing the feature map size by four times while maintaining the number of channels to facilitate subsequent feature extraction. Next, these feature maps enter the main feature extraction structure, which is also the focus of the subsequent validation of LMP-PM. For ease of representation, we divide this portion of the structure into four stages. Let the input image for the first stage be represented as X. Here, we omit the calculations related to batch normalization (BN) and ReLU, and the output of the first stage can be expressed by formula (7) as:

where, Y1X represents the TBP block operation in the first stage, while η denotes the residual module operation. The output of the second stage is similar to that of the first stage, as shown in equation (8):

Compared to the first two stages, the third stage includes an additional SE block. The output of the third stage can be expressed by formula (9):

where, δ represents the Sigmoid activation function in the SE block, while Dense refers to the fully connected layer within the SE block. In the fourth stage, an additional TBP block is added compared to the third stage, and the other components remain similar. Output as shown in formula (10):

where, Y4B3' denotes the operations of the TBP block performed twice. Subsequently, the feature maps outputted from the fourth stage are passed through a 1×1 convolutional layer followed by an Adaptive Pooling layer. Finally, these are flattened and mapped to the output classes by a fully connected layer.

LMNet is a lightweight model derived by applying our proposed Lightweight Multi-Path Pruning Method (LMP-PM) to OMNet, with its structure illustrated in Figure 3B. LMNet significantly reduces the model’s parameter count and computational complexity while maintaining performance superior to OMNet.

The lightweighting of LMNet is primarily achieved through two core strategies: global channel pruning and path expansion. Compared to OMNet, LMNet reduces the global channel count by one-fourth across the entire network. This implies that the channel dimensions of both convolutional and fully connected layers within the network are proportionally scaled down, leading to a substantial reduction in model parameters and FLOPs.

To compensate for potential performance degradation caused by channel pruning and to enhance the model’s feature extraction capabilities, OMNet’s main feature extraction component is expanded into two independent, parallel paths. Each path, in itself, constitutes a pruned and lightweighted OMNet main feature extraction structure. These two paths share the same internal structure, but their parallel processing enables the capture of richer and more diverse feature representations.

Specifically, LMNet’s feature extraction process unfolds as follows: The input image first passes through the same initial downsampling layer as OMNet. Subsequently, the resulting feature maps simultaneously feed into two independent, pruned OMNet main feature extraction structures. These two computational processes are analogous to OMNet’s main feature extraction process detailed in Section 2.4, but with proportionally pruned channel dimensions. The output feature maps from the two parallel paths are then concatenated dimensionally to form a broader feature representation. The final output can be expressed by formula (11):

where, the computation processes for B4 and B4' are similar to those described in Section 2.4. Subsequently, Bout passes through a 1×1 convolutional layer and then an Adaptive Pooling layer before being flattened. Finally, it is mapped to the output classes by a fully connected layer.

We designed a TBP block, and its specific structure is illustrated in Figure 3C. The input image simultaneously enters branches 1, 2, and 3, where corresponding feature extraction tasks are performed before the features are fused. This design facilitates the extraction of multi-scale complex feature information, thereby enhancing the performance of OMNet. Let X denote the input with dimensions (N,C,H,W), where N is the batch size, C is the number of input channels, and H and W represent the height and width of the input feature map, respectively. The output is denoted as Y. The feature map input to branch 1 first passes through a 1×1 convolutional layer, the output is as shown in formula (12):

where, S represents the stride, which is set to 1 by default. The feature map then passes through a second convolutional layer, as shown in formula (13):

After passing through a third convolutional layer, we obtain formula (14):

where Y1,3 undergoes batch normalization, resulting in formula (15):

where μ and σ2 represent the mean and variance of the input, γ and β are the learnable parameters, and ϵ is a small constant to prevent division by zero. Ultimately, the activation mapping produces the result as shown in formula (16):

Similarly, the output of the second branch can be expressed as formula (17):

the output of the third branch is shown in Formula (18):

Finally, the outputs of the three branches are merged through a concatenation operation as shown in Formula (19):

After substituting the numerical values, we obtain the comprehensive output formula of the TBP block as shown in Formula (20):

Our experimental conditions are presented in Table 1. All ablation experiments were conducted under the following conditions: an initial learning rate of 1×10^-4, 100 iterations, the AdamW optimizer (Loshchilov and Hutter, 2017), a weight decay of 5×10^-2, and a cross-entropy loss function. To ensure the comparison reflects the model’s original performance, no pre-training was applied in any of the experiments. In order to validate the effectiveness of LMP-PM and to demonstrate that LMNet possesses strong performance and generalization capabilities, we conducted four relevant experiments. These included verifying the effectiveness of LMP-PM using two excellent open-source agricultural image datasets, testing the efficacy of the TBP block, assessing the lightweight high-performance attributes of LMNet, and finally, performing a visualization validation. To enhance model robustness and prevent overfitting, standard data augmentation techniques were applied to the training data for the PlantVillage dataset. These included Random resized crop (224x224), Random rotation ([-10°, 10°]), and Random horizontal flip (0.5 probability). Images were then normalized to a [0.0, 1.0] range. For the AI Challenger 2018 dataset, no augmentation techniques were applied, as stated in Section 2.2, to specifically evaluate the model’s performance on its inherent data distribution and imbalance.

We employed several commonly used evaluation metrics to assess the performance of LMNet, including accuracy, precision, recall, and the F1 score, as defined by formulas (21-24). We used the number of parameters and FLOPs as indicators of the model’s lightweight characteristics. Additionally, we utilized precision-recall (PR) curves, ROC curves, and confusion matrices to further evaluate LMNet’s classification performance.

where TP denotes the number of true-positive samples, FP denotes the number of false-positive samples, FN denotes the number of false-negative samples, and TN denotes the number of true-negative samples.

To validate the effectiveness of LMP-PM, we designed two sets of experiments, including verification of the lightweight capability of model pruning and assessment of the accuracy enhancement capability through path expansion, thereby comprehensively evaluating the performance of LMP-PM.

To validate the effectiveness of the TBP block, we used LMNet as the baseline model to compare the performance of models with and without the TBP block. The results are presented in Table 7. The model with the TBP block outperformed the model without it on both test sets, achieving a significant increase of 3.88% on the Plant Village dataset and 3.48% on the AI 2018 Challenger dataset. Additionally, all other performance evaluation metrics for the model with the TBP block also surpassed those of the model without it. The accuracy and loss trends during training on the training and validation sets for both models are illustrated in Figure 8. From the figure, it can be observed that the model with the TBP block demonstrates superior convergence speed and performance on both datasets. Notably, in the validation set of the Plant Village dataset, the stability of convergence with the TBP block is significantly better than that without it, effectively proving the validity of the TBP block.

To validate the performance of LMNet, we compared it with six other high-performing models. Among these, ConvNext served as the benchmark model due to its normal complexity and high performance, while the other five models are lightweight models with parameter counts similar to LMNet. The results are presented in Table 8. The experimental results indicate that LMNet achieved a Test accuracy on the Plant Village dataset that is 0.62% higher than the second-ranked Shufflenet_v2_x2_0. Additionally, its Precision is 0.97% higher, F1-score is 0.91% higher, and Recall is 0.82% higher. Notably, LMNet’s Test accuracy surpasses that of the benchmark model Convnext_tiny by 1.12%, while having only one-fifth of the parameters. On the AI 2018 Challenger dataset, LMNet continued to deliver the best performance, exceeding the second-ranked Repvit_m0_9 by 0.92% in Test accuracy, achieving 0.75% higher Precision, 1.38% higher F1-score, and 1.64% higher Recall. Furthermore, LMNet’s Test accuracy also outperformed the benchmark model Convnext_tiny by 5.31%. The changes in Accuracy and loss during training for these seven models on the training set and validation set are shown in Figure 9. From the figure, it is evident that LMNet exhibits superior convergence speed, stability, and overall performance compared to the other models, providing strong evidence for its outstanding capabilities.

All comparative models in Table 9 utilized pre-trained weights. On the Plant Village dataset, LMNet, with only 5.67M parameters, achieved a test accuracy of 99.23%, a score that even surpassed the pre-trained ConvNeXt_tiny (99.11%), despite the latter having a significantly larger parameter count of 28.6M. This comparison strongly demonstrates LMNet’s remarkable advantages in parameter efficiency and performance. Furthermore, on the AI Challenger 2018 dataset, LMNet achieved a test accuracy of 87.27%, ranking first among all models that used pre-trained weights. Moreover, LMNet also exhibited the best performance across key metrics such as Test-acc, Precision, F1-score, and Recall on this dataset. These results conclusively prove that LMNet can demonstrate exceptional generalization ability and leading performance, even without relying on pre-training.

Other results in Table 9 indicate that models like ConvNeXt_tiny, RegNetX_400MF, and PVT_Tiny showed significant performance improvements after loading pre-trained weights. For instance, ConvNeXt_tiny’s test accuracy on the Plant Village dataset increased from 98.11% (as shown in Table 8) to 99.11%, and on the AI Challenger 2018 dataset, it saw a substantial increase from 81.96% to 86.24%. RegNetX_400MF’s test accuracy on the Plant Village dataset also rose from 97.06% to 99.03%. Similarly, PVT_Tiny’s test accuracy on the Plant Village dataset improved from 96.89% to 99.04%. This highlights the critical role of pre-training in enhancing the performance of these general-purpose models on target tasks.

However, not all models exhibited improvement after loading pre-trained weights. For example, Mobilenet_v3_large and EfficientNetB0 even showed a slight performance decrease on the AI Challenger 2018 dataset, which can be attributed to the domain discrepancy between the pre-training task and the downstream task. LMNet’s ability to achieve such excellent results without pre-training, especially surpassing SOTA models that rely on ImageNet pre-training, strongly suggests its architectural advantages in learning specific agricultural features. LMNet’s design is more focused on extracting unique, fine-grained visual patterns specific to plant diseases and agricultural scenes, such as subtle variations in leaf texture, the shape and color characteristics of lesions, and nuanced differences in crop growth stages.

Unlike general SOTA models (e.g., ConvNeXt_tiny), which are typically pre-trained on large-scale, diverse general datasets like ImageNet, the features they learn may be more geared towards general object recognition. When these general features are transferred to highly specialized agricultural domains, a “domain mismatch” issue can arise, leading to reduced effectiveness of some general features for agricultural tasks, or even introducing noise. LMNet, through its specially designed modules and connection mechanisms, likely avoids the interference of general features, directly and efficiently learning and encoding these domain-specific, highly discriminative features from agricultural data. This “from-scratch,” targeted learning approach enables LMNet to better adapt to the characteristics of agricultural images, thereby demonstrating excellent generalization ability and performance even without the need for external large-scale pre-training data.

Although LMNet achieved the best performance on both the Plant Village dataset and the AI Challenger 2018 dataset, the performance gap between these two datasets reached as much as 11.96%. To investigate the reasons behind this discrepancy, we utilized Matplotlib to plot the ROC curves, PR curves, and confusion matrices for LMNet on these two datasets, visualizing its performance and classification capabilities across different categories. The results are presented in Figures 9–11. From Figure 9, it is evident that the ROC curves for LMNet on both datasets are very close to the top left corner, indicating a high true positive rate and a low false positive rate. Furthermore, the AUC values approach or equal 1, demonstrating LMNet’s robust ability to identify positive samples. In Figure 10, the PR curve for LMNet on the Plant Village dataset also approaches the top right corner, showing high precision and recall. However, the multiple PR curves for LMNet on the AI Challenger 2018 dataset do not converge towards the top right corner. Notably, labels 15, 39, 48, and 49 exhibit substantial deviations, suggesting that certain categories in the AI Challenger 2018 dataset have very few positive samples. This scarcity likely contributes to the poor performance of most classification models on this dataset. The confusion matrix depicted in Figure 11 clearly illustrates the number of correctly classified samples for all categories by LMNet on both datasets. LMNet successfully identified the majority of disease images in the Plant Village dataset; however, it made classification errors for several categories in the AI Challenger 2018 dataset, including labels 18, 19, 54, and 55. This indicates that specific categories require attention in future model optimization efforts.

Finally, we used Grad-CAM (Selvaraju et al., [[NoYear]]) to visualize the feature extraction process of LMNet and the class activation maps of some exemplary images, as shown in Figures 12, 13. In Figure 12, using the first 15 channels as an example, we observe that different models have similar extraction effects on the shallow features (such as shape and color) of diseased images during the feature extraction process. However, in the extraction of deeper features for plant leaf diseases, the model without the TBP block module fails to extract many abstract features due to its weaker multi-scale extraction capability. Both LMNet and OMNet can extract more in-depth abstract features. However, while OMNet extracts a large number of abstract features, some of its channel feature maps show overly smooth edges, such as channels 4, 5, 9, and 10. This could be due to a slight overfitting caused by OMNet’s higher complexity. When outputting the disease category, OMNet has the highest score for the correct category but fails to sufficiently suppress incorrect categories. In contrast, LMNet shows the most concentrated correct category scores and the most suppressed incorrect category scores, indicating its superior feature extraction capabilities. In Figure 13, we can see that LMNet identifies a larger area and with more precise localization of leaf diseases compared to ConvNeXt-Tiny and Repvit_m0_9. This suggests that LMNet performs better in identifying plant leaf diseases.