CNN5Conv2FC: A Resource‐Aware Deep Learning Framework for Real‐Time Mango Leaf Disease Recognition

Mango cultivation plays a vital role in the agricultural economy of tropical and subtropical regions, serving as a key source of food security and livelihood for millions of farmers. Despite its importance, mango production faces significant challenges from leaf diseases such as anthracnose, powdery mildew, and bacterial canker, which severely impact crop yield and quality. In countries like Bangladesh, traditional disease diagnostic methods remain inefficient and largely inaccessible, particularly in rural areas where agricultural expertise and laboratory facilities are limited. To address these challenges, this study proposes a lightweight convolutional neural network (CNN) architecture, comprising five convolutional layers and two fully connected (2FC) layers (CNN5Conv2FC), designed specifically for automated classification of mango leaf diseases. The proposed model incorporates five convolutional layers that progressively increase filter sizes for improved feature extraction, followed by 2FC layers with dropout, along with batch normalization and max pooling to ensure robust generalization. Together, these design choices enable efficient feature learning while maintaining minimal computational complexity. Experimental evaluation demonstrates that CNN5Conv2FC achieves a classification accuracy of 97.74%, with precision, recall, and F1‐score values of 97.78%, 97.74%, and 97.73%, respectively. Additionally, the model exhibits the lowest total inference time of 1.3447 s on the 398‐image test set and an average batched inference time of only 0.003379 s per image, making it highly suitable for real‐time deployment in resource‐constrained agricultural environments. To establish reliability and transparency beyond headline accuracy, the model is further assessed through a confusion matrix, per‐class receiver operating characteristic (ROC) area under the curve (AUC) (≥0.99), bootstrap confidence intervals, corruption‐based robustness tests, hardware‐agnostic complexity metrics (FLOPs and model size), and Grad‐CAM and Grad‐CAM++ interpretability.