Machine Vision–Based Classification of Rare Fruits in Bangladesh Using Transfer Learning and Custom CNN Models

Abstract

This study presents a deep learning framework for recognizing six rare Bangladeshi fruits using image classification. A dataset of 1,800 real-world images was created, and a lightweight Custom CNN was compared with transfer-learning models including MobileNetV2, InceptionV3, ResNet-50, and DenseNet-121. Experimental results show that the proposed Custom CNN achieved the best performance with 97.40% accuracy, along with superior ROC–AUC and PR–AUC scores. The findings demonstrate that domain-specific lightweight CNN architectures can outperform deeper pre-trained models while remaining computationally efficient. The proposed system has potential applications in digital agriculture, biodiversity conservation, and educational awareness related to Bangladesh’s rare fruit heritage.