A Lightweight Family of CNN-MLP Framework for Apple Leaf Disease Classification on Augmentation-Enhanced Heterogeneous Dataset

Accurate, computationally efficient classification of apple leaf diseases is essential to improving orchard productivity. Despite advances in deep learning, many existing models rely on intensive architectures that limit deployment in resource-constrained environments. This study proposes a lightweight CNN-MLP framework optimized for disease classification under heterogeneous visual conditions. Pre-trained ResNet-۱۸ and EfficientNet-B۰ backbones were utilized for feature extraction, with their original classification heads replaced by a task-specific, two-layer multilayer perceptron (MLP) head. This architectural adaptation, incorporating a ۲۵۶-neuron hidden layer and dropout regularization, aims to enhance learning of non-linear boundaries in diverse datasets. An experimental evaluation on the ۹-class AL۹EE dataset, conducted across five independent runs, demonstrated that the EfficientNet-B۰-MLP model achieved ۹۸.۴۲ ± ۰.۲۶% accuracy and a macro F۱-score of ۹۸.۳۳ ± ۰.۳۴%, outperforming the ResNet-۱۸-MLP variant. The lower standard deviations indicate improved training stability and reproducibility. By balancing predictive performance with reduced computational complexity, this framework provides a practical solution for edge-oriented agricultural applications, offering an effective pathway for intelligent disease monitoring in smart farming.

کلیدواژه ها:

Apple leaf disease classification ، hybrid deep learning ، EfficientNet-B۰ ، ResNet-۱۸ ، Lightweight deep learning Heterogeneous visual conditions