| dc.description.abstract | The accurate detection and classification of brinjal (eggplant) leaf diseases are crucial for sustainable agriculture, as early diagnosis can prevent crop loss and improve yield quality. Traditional manual identification is often time-consuming, error-prone, and reliant on expert knowledge, underscoring the need for automated and intelligent solutions. This study presents a comparative evaluation of three pre-trained convolutional neural network models MobileNetV2, DenseNet121, and EfficientNetB4 applied to brinjal leaf disease detection using transfer learning, fine-tuning, data augmentation, and class balancing strategies. The dataset, encompassing multiple disease categories including Healthy, Insect Pest, Leaf Spot, Mosaic Virus, Small Leaf, White Mold, and Wilt, was preprocessed and augmented to ensure balanced representation. Experimental results indicate that DenseNet121 achieved 96.74% accuracy, demonstrating the benefits of dense feature connectivity, while EfficientNetB4 attained the highest accuracy of 98.57%, leveraging compound scaling and larger input resolution at the cost of increased computational demand. MobileNetV2 achieved 98.29% accuracy with only 3.5 million parameters, offering superior inference speed suitable for real-time and mobile applications. Regularization techniques such as dropout, L2 weight decay, and class weighting were integrated to mitigate overfitting and improve minority class recall, enhancing model robustness. This work highlights the trade-offs between accuracy, computational efficiency, and deployment feasibility, providing a practical framework for precision agriculture. The findings suggest MobileNetV2 as an optimal choice for field-level applications, EfficientNetB4 for high-performance scenarios, and DenseNet121 as a balanced alternative, while future work will focus on scaling datasets, mobile deployment, and integrating explainable AI to improve trust and usability. | en_US |
