WEIGHT HIPPOPOTAMUS OPTIMIZATION ALGORITHM MULTI THRESHOLD SEGMENTATION (WHOAMTS) AND MULTI-SCALE ATTENTION MULTI-AXIS VISION TRANSFORMER (MAMViT) FOR TOMATO LEAF DISEASES DETECTION

Abstract

 The rapid integration of artificial intelligence (AI) and computer vision into precision agriculture has revolutionized disease monitoring and crop management by enabling automated, accurate, and early detection of plant health conditions. In particular, tomato cultivation, which holds substantial economic importance worldwide, faces major productivity losses due to various foliar diseases. Early diagnosis of these diseases is crucial to prevent infection spread and minimize yield reduction. To tackle this challenge, this paper presents a comprehensive hybrid framework combining the Weight Hippopotamus Optimization Algorithm Multi-Threshold Segmentation (WHOAMTS) and the Multi-Scale Attention Multi-Axis Vision Transformer (MAMViT) for efficient tomato leaf disease detection. Initially, a Lagrange Conditional Generative Adversarial Network (LCGAN) is employed to pre-process and denoise tomato leaf images, producing high-quality data for subsequent segmentation. The LCGAN follows an encoder–decoder architecture guided by a Lagrangian constraint to minimize reconstruction error, effectively reducing illumination distortion and background noise. This enhances the visibility of diseased regions and simplifies the extraction of discriminative features for classification. After denoising, the WHOAMTS algorithm performs multi-threshold segmentation by determining optimal thresholds inspired by the natural behaviour of hippopotamuses— MAMViT such as defense mechanisms, territorial movement, and adaptive escape strategies. These behaviours are mathematically modelled to improve the global search balance and segmentation accuracy across multiple grayscale levels, ensuring precise region separation between infected and healthy tissues. Following segmentation, the Inception-V3 model is employed to extract abundant disease-specific features from the segmented regions, enabling rich feature representation for classification. The extracted features are then processed by the framework, which integrates convolutional feature encoding with multi-axis self-attention to capture both fine local structures and global dependencies. This dual-domain representation enables the network to recognize subtle visual cues in various tomato leaf diseases, including Bacterial Spot, Early Blight, Leaf Mold, Septoria Leaf Spot, Leaf Curl Virus, and Healthy categories. Experimental analysis conducted on a real-time tomato leaf dataset confirms that the proposed LCGAN–WHOAMTS–Inception-V3–MAMViT pipeline achieves superior detection performance over conventional methods, with significant improvements in DSC, JSI, precision, recall, F-measure, and overall accuracy. The findings demonstrate the potential of the proposed framework as a robust and intelligent diagnostic system for sustainable and smart agricultural applications.