A Comparative Study of Different CNN Architectures for Real-World Image Classification in Bangladesh

Convolutional Neural Networks (CNNs) are widely used for image classification, yet their performance strongly depends on dataset complexity and deployment constraints. This study presents a comparative evaluation of custom-designed CNN architectures and popular pre-trained models on five real-world image datasets from Bangladesh, spanning agricultural and infrastructural applications. The tasks include mango variety classification (15 classes), paddy disease clas-sification (35 classes), and three binary classification problems: road damage, footpath encroachment, and auto-rickshaw detection. In addition to task-specific CNNs, VGG16 and ResNet50 are evaluated using fixed feature extraction and transfer learning strategies. The results show that transfer learning, particularly with ResNet50, achieves the highest accuracy on complex multi-class datasets, while custom CNNs deliver competitive performance on binary tasks with sub-stantially lower computational cost. These findings emphasize the trade-off between accuracy and efficiency and highlight the importance of selecting model architectures based on dataset characteristics and deployment requirements.