Several traditional methods have been proposed for segmenting plant leaf images. Threshold-based techniques, such as fuzzy C-means algorithms (Liu, 2012), are commonly used to iteratively determine the optimal threshold for leaf image segmentation. Histogram-based thresholding methods, including bimodal histograms and Otsu’s Thresholding Method (Kalaivani et al., 2020; Fang et al., 2021b), have also been employed for segmenting leaf images. However, these threshold-based methods often struggle with complex images. Region-based approaches, such as the region-based level set method (Fang et al., 2021a), region growing methods (Jothiaruna et al., 2021), and wavelet methods (Xiong et al., 2020), have shown high accuracy and fast processing speeds for plant leaf segmentation. While these methods yield satisfactory results to some extent, their effectiveness is heavily dependent on image features, which limits their broader applicability. Clustering-based methods, such as fuzzy k-means clustering (Tian et al., 2019), have been used to determine cluster centers for leaf segmentation. However, these methods often struggle with local optima, leading to lower segmentation accuracy.

Deep learning techniques have revolutionized the field of image segmentation, with convolutional neural networks (CNNs) playing a pivotal role. The introduction of fully convolutional networks (FCN) by Long et al. (2015) marked a significant milestone, replacing traditional fully connected layers with specialized convolutional layers tailored for segmentation tasks. Building on this, Ronneberger et al. (2015) proposed the U-Net architecture, which employs an encoder-decoder structure with skip connections to fuse low-level and high-level features. U-Net and its variants, such as R2U-Net (Alom et al., 2018) and BIONet (Xiang et al., 2020), have shown strong performance in segmentation, particularly for medical and agricultural applications. However, despite their success, CNN-based methods often face challenges in extracting detailed boundary information, especially in complex and varied environments.

To address these limitations, researchers have incorporated attention mechanisms into CNNs (Oktay et al., 2018; Zhou et al., 2019). For example, the squeeze-and-excitation network (SE-Net) (Hu et al., 2018) uses channel-wise attention to enhance global feature representation, while the attention-guided network (Li et al., 2019a) focuses on suppressing irrelevant background information. A parallel reverse attention network (PraNet) (Fan et al., 2020) introduced a reverse attention block to build relationships among object regions and boundaries. Despite their improvements, these models still struggle with precise boundary delineation, especially in complex segmentation tasks such as plant disease recognition.

Transformer-based models, originally designed for natural language processing (Devlin et al., 2018), have been adapted for computer vision tasks, including image segmentation. These models use self-attention mechanisms to capture long-range dependencies in images, improving segmentation accuracy for global features. For instance, TransUNet (Chen et al., 2021) combines the U-Net architecture with transformers to leverage high-level informative features for improved performance. Fang et al. (2023) proposed BAF-Net, a network combining CNNs and Swin Transformers for plant leaf segmentation. It utilizes MSFF and FSFF branches, enhanced by an adaptive bidirectional attention module, to capture comprehensive features. Dai et al. (2024) introduce AISOA-SSformer, a Transformer-based segmentation method for rice leaf disease detection. By integrating sparse global updates, feature attention, and optimized algorithms, it achieves high accuracy, aiding modern farming. However, transformer models often focus primarily on global context and struggle with capturing fine-grained details, such as object boundaries.

Recently, multi-layer perceptron (MLP)-based models have gained attention as a viable alternative to CNNs and transformers for image segmentation. The MLP-Mixer (Tolstikhin et al., 2021) demonstrated that MLPs could replace self-attention mechanisms in image processing, achieving competitive performance in tasks like image classification. This idea was further explored in the Visual Transformer (ViT) (Melas-Kyriazi, 2021), where MLPs replaced the attention layers, showing that MLP-based networks could achieve similar results to CNNs and transformers in recognition tasks.

In segmentation tasks, MLP-based models like RepMLPNet (Ding et al., 2022) and MAXIM (Tu et al., 2022) have been shown to effectively replace self-attention mechanisms while maintaining high accuracy. These models utilize fully connected layers to capture both local and global context information, making them suitable for complex image segmentation tasks. Additionally, MLPs have been integrated with CNN architectures to form hybrid models that combine the benefits of both approaches. For instance, Valanarasu and Patel (2022) introduced UNeXt, a convolutional MLP-based network with a U-shaped architecture, comprising three convolution blocks and two tokenized MLP blocks for global information capture and pixel-wise classification. Similarly, the CM-MLP framework (Lv et al., 2022) integrates multi-scale feature interaction (MSFI) and axial context encoder (ACE) blocks, enhancing local information integration and establishing edge relations between foreground and background regions.

Inspired by the strengths of MLP-based models and transformers, our approach, the Adaptive Multi-Scale MLP (AMS-MLP), combines the benefits of both architectures to address these challenges. The AMS-MLP model integrates multi-path aggregation and multi-scale context relation modules, enabling dynamic fusion of global and local features for accurate segmentation, especially in complex backgrounds. To further highlight the novelty of our work, we provide a comprehensive comparison with existing leaf segmentation methods in Table 1 . The table focuses on summarizing the Key Features, Strengths, and Limitations of existing methods, while explicitly outlining How AMS-MLP Differs/Improves over these approaches. Unlike previous approaches, AMS-MLP uniquely leverages adaptive multi-scale feature fusion and context-aware modeling, which significantly improves segmentation accuracy in challenging scenarios. This comparative analysis underscores the advancements of our method and its distinct contributions to the field of leaf segmentation.

The pepper leaf image datasets utilized in this study were sourced from a private repository maintained by the Nanchang Academy of Agricultural Sciences. These datasets were collected from their farm located in Nanchang city, Jiangxi Province, China, specifically between August 12 and 13, 2022, using multi-view photography techniques. The camera used for image acquisition was equipped with an F5.6 lens and an EF-S 18-135mm f/3.5-5.6 IS USM microlens manufactured by Canon Company (Japan). During the data collection process, the camera was positioned at a height of 10-50 cm above the leaves to ensure high-resolution images. All images were captured after careful focusing, and the camera remained stationary during shooting to eliminate any motion blur or distortion caused by movement.

To ensure the comprehensiveness and practicality of the data, stringent inclusion criteria were established, encompassing various instances of pepper leaf diseases, including healthy leaves and those affected by viral infections. Specifically, the dataset includes images of leaves severely impacted by common diseases such as early blight, brown spot disease, and leaf mold, along with healthy pepper leaves (HPL) and viral diseases (VD) to enrich the diversity of the dataset. This hybrid database serves as a valuable resource for researching and developing methodologies related to pepper leaf segmentation and disease classification in agricultural research.

To further evaluate the effectiveness of the proposed model in segmenting actual pepper leaves, four distinct datasets were meticulously constructed: Early Blight Dataset (EBD), Brown Spot Dataset (BSD), Leaf Mold Dataset (LMD), and Mixed Leaf Dataset (MLD). These datasets were manually annotated using the open-source tool LabelMe, assigning intensity values of 1 to foreground regions and 0 to background regions. During the data processing phase, we conducted a meticulous statistical analysis to ensure the representativeness and balance of the data across various disease categories and leaf conditions. This analysis involved calculating the distribution of images among different disease categories, conducting rigorous checks on the completeness and quality of annotations, and verifying the accuracy and reliability of the data. As shown in Table 2 , the datasets for EBD, BSD,LMD, and MLD comprised 1190, 1384, 1385, and 6613 images, respectively. Notably, the MLD dataset integrates image data from EBD, BSD, LMD, as well as healthy pepper leaves and viral diseases, with 1353 images of healthy leaves and 1301 images of viral leaves. Figure 1 shows several representative examples from the EBD, BSD, MLD and HPL, respectively. The statistical analysis confirmed the balanced nature of the MLD dataset, and all data were comprehensively annotated. Furthermore, to facilitate a comprehensive evaluation, each dataset was divided into training (70%), validation (10%), and testing (20%) subsets. In the experiment, we standardized the image size for each dataset to 512×512 pixels, facilitating consistent processing and analysis across all datasets.

In this section, we will provide an overview of the AMS-MLP model and discuss the incorporation of three key modules within the encoder-decoder architecture. These modules consist of the adaptive multi-scale MLP module, the multi-scale context relation decoder module, and the multi-path aggregation mask module. Additionally, we will present the loss function utilized in the model. By integrating these modules and utilizing an appropriate loss function, the AMS-MLP model demonstrates improved performance in image segmentation tasks.

The proposed AMS-MLP network was implemented on a specific hardware configuration consisting of a 12th Gen Intel(R) Core(TM) i7-12700K 3.60 GHz processor and an NVIDIA GeForce RTX 3090 40 GB GPU with 32 GB of RAM. The operating system employed was Windows 11, and the Conda environment was utilized to ensure a consistent software environment for the execution.

For the parameter settings of the AMS-MLP network, we employed a batch size of 2 and trained the model for a total of 60 epochs. The optimization process was performed using stochastic gradient descent (SGD) with an initial learning rate of 0.001, and the learning rate was adjusted according to the learning rate schedule. During training, these parameter choices facilitated effective convergence of the AMS-MLP network. To assess the model’s performance, we compared it against several state-of-the-art (SOTA) models using various performance metrics. Additionally, ablation studies were conducted to evaluate the impact of different network components, such as the choice of activation functions and the depth of layers, on the overall performance. In our experiments, we also tested other values for batch size, learning rate, and epoch count, but found that the chosen configuration yielded the best performance in terms of both training stability and testing accuracy.

The proposed AMS-MLP network involves two loss functions to optimize the predicted result and the ground truth (GT), including the binary cross entropy (BCE) L b and the Dice L d . The two loss functions are defined as Equations 24, 25:

where f denotes the input predicted result, and g denotes the corresponding ground truth label.

Therefore, our final loss L l o s s can be expressed as Equation 26:

To rigorously evaluate the performance of the proposed method and other compared methods, six metrics are employed as evaluation criteria: accuracy, recall, precision, specificity, F1-score, and intersection over union (IoU). Here’s a detailed breakdown of these metrics. These metrics are defined as follows:

To evaluate the segmentation performance of the AMS-MLP model, we conducted a comparative study against state-of-the-art (SOTA) models using three distinct leaf datasets: EBD, BSD, and MLD. The models included FCN-VGG16, U-Net (Ronneberger et al., 2015), attention U-Net (AttU-Net) (Oktay et al., 2018), UNet++ (Zhou et al., 2019), UNeXt (Valanarasu and Patel, 2022), and CM-MLP model (Lv et al., 2022). To ensure a fair and comprehensive comparison, all models were trained, validated, and tested on the same three datasets. By maintaining consistency across the training, validation, and test datasets, we aim to eliminate any potential bias or variation that may affect the results and evaluate the segmentation performance exclusively on the test dataset.

Table 3 and Figure 7 present the segmentation results of the AMS-MLP model compared to seven other models on the EBD dataset, evaluated across six metrics: accuracy, recall, precision, mean Intersection over Union (mIoU), and F1-score. Our model achieves the highest scores across five of these metrics, notably achieving 97.39% in mIoU and 98.29% in F1-score, surpassing FCN by 0.35% and 0.29%, respectively. These superior metrics reflect the AMS-MLP model’s ability to leverage the GMS-MLP and LMS-MLP modules effectively. The GMS-MLP module captures global context, enabling robust feature extraction across the entire image, while the LMS-MLP module enhances local detail recognition within specific leaf structures. This dual-stream approach ensures comprehensive information integration, leading to enhanced segmentation accuracy. Compared to U-Net, our model demonstrates significant improvements with increases of 9.92% in mIoU, 0.18% in F1-score, and 0.29% in recall. This enhancement can be attributed to the AMS-MLP’s capacity to combine both global and local features effectively, thereby improving boundary delineation and reducing segmentation errors. Furthermore, compared to other semantic segmentation models, our approach achieves the highest scores in accuracy, recall, precision, mIoU, and F1-score, underscoring its robust performance across multiple evaluation criteria. Qualitative examples in Figure 8 highlight the models’ ability to locate object regions accurately, with our proposed model notably delineating pepper leaf boundaries with precision. This visual evidence further substantiates the effectiveness and superiority of the AMS-MLP model in leaf segmentation tasks.

Additionally, for a comprehensive evaluation of training performance, we conducted a comparative analysis of the AMS-MLP network against FCN-based, U-Net-based, and additional MLP-based models, utilizing the BSD and MLD datasets as benchmarks. The results of this comparison are presented in Tables 4 , 5 , and Figures 9 , 10 . Our method consistently outperformed other models across five evaluation metrics. This improvement can be attributed to the integration of GMS-MLP and LMS-MLP modules, enabling the extraction of both global and local information crucial for enhancing segmentation accuracy. The FCN-16s model’s superior performance over U-Net is attributed to its use of pretrained VGG16, which enriches feature extraction and representation within the encoder. Similarly, the CM-MLP model, by leveraging MLP instead of traditional attention mechanisms, achieves superior segmentation results by effectively considering pixel relationships.

Figure 11 , 12 provide visual insights into segmentation outputs, revealing that U-Net and UNet++ models exhibit certain false positive regions in lesion segmentation. In contrast, AttU-Net shows improved performance over U-Net, while the AMS-MLP model closely approximates ground truth, demonstrating precise extraction of pepper leaf boundaries and reduced false positive regions. This performance superiority is facilitated by the MCRD module and the utilization of GMS-MLP and LMS-MLP auxiliary streams, which facilitate effective cascaded contraction and expansion processes within the network.

Despite these advancements, our study acknowledges areas for further improvement, particularly in optimizing computational efficiency for real-time applications and scaling the model to larger datasets. Specifically, we recognize that, in terms of model size, our proposed AMS-MLP does not demonstrate a significant advantage compared to other models, such as CM-MLP, FCN-16s, U-Net, AttU-Net, and UNet++, as evidenced by our analysis of memory footprint and storage requirements. Nevertheless, we believe our method retains value in other critical aspects, such as potentially offering higher accuracy or efficiency in specific data processing scenarios. Future research will focus on refining architectural designs, exploring advanced training strategies, and actively working on optimizing our method to address these challenges and further elevate segmentation performance. Our ongoing efforts include strategies to reduce resource consumption and enhance overall performance, aiming to better meet practical application needs.

In this section, we conducted a comprehensive ablation study to systematically evaluate the impact of individual modules on segmentation performance using the MLD dataset. Our baseline model was derived from the BU-Net architecture, with a modified channel configuration to reduce model complexity. To assess the contribution of each module, we adopted a phased integration approach. Initially, we introduced the BAM-MLP model by incorporating the AM-MLP module into the BU-Net architecture. The AM-MLP module enhances the network’s ability to capture global context information, which improves the focus on informative regions, thereby boosting segmentation performance. Subsequently, we integrated the MPAM module into the encoder, specifically at the fifth layer, resulting in the BMAM-MLP model. The MPAM module is crucial for generating precise masks, refining the segmentation process by better delineating object boundaries. Further, to explore the synergistic effects of incorporating multiple modules, we tested the BMRD-MLP model, which integrates both the AM-MLP and MCRD modules into the BU-Net architecture. The MCRD module enhances the network’s ability to preserve boundary details, improving segmentation accuracy by focusing on fine-grained features. Finally, we developed the AMS-MLP model by progressively combining the AM-MLP, MPAM, and MCRD modules. This multi-module integration demonstrated superior performance across multiple evaluation metrics, illustrating the complementary effects of these modules in enhancing segmentation accuracy. Through this incremental approach, we were able to systematically assess the contribution of each module. As shown in Figure 13 , the results clearly demonstrate the positive impact of each module on enhancing the overall segmentation performance.

Table 6 presents the detailed results of all ablation experiments. As illustrated in Table 5 , our study commenced with the BU-Net model, which achieved the following performance metrics: 97.65% accuracy, 93.92% recall, 79.11% precision, 97.96% specificity, 97.74% mIoU, and 85.88% F1-score. To improve upon this baseline, we incorporated the AM-MLP module into the BU-Net architecture, resulting in the BAM-MLP model. A comparative analysis revealed that BAM-MLP outperformed BU-Net across all metrics, achieving 98.37% accuracy, 96.70% recall, 84.20% precision, 98.51% mIoU, and 90.02% F1-score—improvements of 0.72%, 2.78%, 0.55%, 0.32%, and 4.14%, respectively. Building on these results, we investigated the synergistic effects of adding more modules to the BAM-MLP model. We first integrated the MPAM module, leading to the BMAM-MLP model, and then included the MCRD module, resulting in the BMRD-MLP model. Both modifications brought about substantial improvements in segmentation performance, with noticeable gains in accuracy, recall, precision, specificity, and F1-score compared to the BAM-MLP model. Lastly, we constructed the AMS-MLP network by incorporating both the MPAM and MCRD modules into the BAM-MLP architecture. The MPAM and MCRD modules effectively harnessed multi-scale features, preserving boundary details and further enhancing segmentation performance. The integration of these three modules—AMSS-MLP, MPAM, and MCRD—yielded the best overall performance, as demonstrated by a thorough analysis of the combined metrics. This analysis clearly highlights the distinct contributions of each module when applied to the MLD dataset.

The primary goal of this study is to develop a precise pepper leaf segmentation model to support ground-based mobile disease recognition systems. The proposed Adaptive Multi-Scale MLP (AMS-MLP) model is specifically designed to segment diseased leaves from images captured under various pure backgrounds (e.g., palm, ground, or desktop), which are commonly used in ground-based data collection scenarios. By accurately extracting leaves from complex backgrounds, the segmented leaves can be fed into disease recognition models, significantly improving their accuracy and robustness.

The AMS-MLP model is optimized for deployment on mobile devices (e.g., smartphones), making it highly accessible and practical for small-scale farmers. In real-world applications, farmers or agricultural workers can use a smartphone to capture images of pepper leaves placed on different pure backgrounds in the field. The AMS-MLP model will then precisely segment the leaves, enabling subsequent disease recognition models to perform more effectively. This approach is particularly beneficial for farmers who rely on mobile devices for crop monitoring and disease detection, as it provides a cost-effective and user-friendly solution.

While our current focus is on ground-based mobile applications, the AMS-MLP model can also be adapted for other platforms, such as UAVs (drones) or robotic systems, in future work. For example, integrating the model into drones could enable large-scale monitoring of pepper fields, while embedding it into agricultural robots could support automated disease detection and precision farming. However, the immediate application of our research is to enhance the performance of mobile-based disease recognition systems by improving leaf segmentation accuracy under varying backgrounds.