Simple Summary
Early and accurate identification of cattle lumpy skin disease is essential for protecting animal health and reducing production losses. However, subtle skin lesions and visual similarities between disease categories make automated classification challenging. In this study, a Lesion-Aware Global–Local Network (LANet) is proposed to address these challenges by combining information from the entire image with attention to informative local regions. By integrating complementary global and local visual features, LANet offers a promising approach to accurate and efficient automated classification of cattle diseases.
Abstract
Lumpy Skin Disease (LSD) threatens cattle health and productivity, imposing economic burdens on livestock production and highlighting the need for reliable image-based classification methods. Although deep learning has demonstrated strong potential for disease-related image classification, distinguishing localized and fine-grained visual patterns remains challenging. In this study, we propose LANet, a Lesion-Aware Global–Local Network for multi-class cattle LSD classification. The proposed framework employs a pretrained ResNet50 backbone followed by two complementary feature branches: a global branch based on adaptive average pooling and a local branch that learns a spatial attention mask to emphasize informative feature regions before adaptive max pooling. The resulting global and local representations are integrated through a feature interaction module and subsequently refined using adaptive feature gating before classification. Experiments on an 8012-image, five-class cattle disease dataset demonstrate that LANet achieves 98.93% accuracy, 97.67% macro-precision, 98.23% macro-F1, and 98.43% Cohen’s kappa against the dataset annotations. The model has 59.18 M parameters, requires 6.621 GFLOPs under standard inference, and achieves an inference latency of 0.533 ms per image with a throughput of 1875.30 images/s. These results demonstrate the effectiveness of LANet’s explicit global–local feature-learning and integration strategy, which combines holistic feature aggregation, spatially guided local feature extraction, feature interaction, and adaptive gating for multi-class classification of cattle disease images based on the available dataset annotations.