Attention-Driven Dual Optimization Framework for Accurate and Generalizable Maize Leaf Disease Classification

Abstract

This paper, proposes a Dual-Optimization Convolutional Neural Network to improve the accuracy and generalization of maize leaf disease classification. To address the constraints of predefined CNN architectures and single-algorithm hyperparameter consistence optimization, we combine two meta-heuristics optimizers with an architectural attention layer. In the proposed dual-optimization framework, PSO determines the number of convolutional filters in a dynamic manner, and BO finetunes the size of dense layers and global learning rate. The intermediate hybrid CNN structure is then instantiated with the best found FCSs and two other models are inspected: the hyper parameter hybrid model and the improved ancestral squeeze and excitation-based hybrid model that introduces modified SE attention blocks for concentrating on more discriminative, lesionrelated zones. All models were trained with focal loss, dynamic L2 regularization, and Learning Rate Scheduler to obtain solid and stable class imbalance handling. Competitive experiments over the test set show that the hybrid model leads to higher classification accuracy and generalization, which confirms that combined optimization on hyperparameters and attention-driven learning directly supports establishing a cutting-edge diagnostic model in our proposed co-design strategy. The proposed model achieved an accuracy of approximately 89%, outperforming the baseline CNN, PSO, and BO models, thus validating the effectiveness of the dual optimization framework