DenseNet Model with Attention Mechanisms for Robust Date Fruit Image Classification
Esraa Hassan, Sarah Abu Ghazalah, Nora El-Rashidy, Tarek Abd El‐Hafeez, Mahmoud Y. Shams
Abstract
Dates in Saudi Arabia hold immense cultural, economic, and nutritional importance, being a staple food and a symbol of heritage. Saudi Arabia produces approximately 1.5 million metric tons of dates annually, accounting for nearly 17% of global date production, underscoring its pivotal role in the global market. Precise and automated classification of fruits is a crucial aspect of modern agriculture, yet it remains a challenging endeavor due to the diverse appearances of fruits. The classification of date fruit varieties presents additional complexities, given variations in size, shape, and texture, making it a critical focus for technological advancements. Dates are highly nutritious, providing approximately 277 cal per 100 g and serving as an excellent source of dietary fiber, natural sugars, and energy. Their substantial nutritional value makes them indispensable in addressing food security challenges and promoting global health benefits. In this paper, we introduce a novel DenseNet-based model augmented with attention mechanisms and optimized using the Nadam algorithm. Unlike traditional DenseNet variants, the model integrates attention mechanisms to enhance focus on pertinent image features, thereby improving classification accuracy under challenging conditions. To evaluate its efficacy, the model was benchmarked against several state-of-the-art deep learning architectures, including DenseNet with Adam optimization, EfficientNet, GoogleNet, HRNet, MobileNet, and VGG, optimized with both Adam and Nadam algorithms. The proposed model achieved outstanding performance metrics, including 98.05% accuracy, 98.00% precision, 97.04% recall, and a 98.32% F1-score.