Enhancement of Ship Classification Accuracy with Balanced Data and Convolutional Neural Networks

This paper introduces a framework for enhancing ship classification by balancing a dataset of $\mathbf{8, 9 3 2}$ images ($\mathbf{6, 2 5 2}$ for train, $\mathbf{2, 6 8 0}$ for test for Cargo, Military, Carrier, Cruise, and Tankers) to $\mathbf{1 5, 0 0 0}$ images ($\mathbf{3, 0 0 0}$ per category) through data augmentation. Four CNN architectures are considered: Custom CNN, VGG16, ResNet50, and MobileNet. The models are assessed using accuracy, precision, recall, and F1-score after training on $224 \times 224$ RGB inputs. MobileNet achieved 71.40%, whereas VGG16 had the best accuracy of $\mathbf{9 3. 0 3 \%}$ as a consequence of issues with Cargo and Tankers. All the models are implemented for real-time classification through a Streamlit web application, providing a set of predictions (for instance, a cruise ship identified as Cruise by Custom CNN by $\mathbf{3 2. 7 5 \%}$, Carrier by MobileNet by $\mathbf{9 9. 2 7 \%). ~ T h e ~ p a p e r ~ e m p h a s i z e s ~ m u l t i - m o d e l ~ a n a l y s i s ~ a s ~ w e l l ~ a s ~}$ balanced datasets; further research will concentrate on hybrid methods as well as underperforming classes.