The proliferation of digital transactions and the growing need for secure identity verification systems have underscored the importance of efficient and accurate ID card classification methods. This research addresses the challenge of accurately classifying Fiji-based identification cards using Convolutional Neural Networks (CNNs). With a private dataset of over 11,000 images encompassing various identification card types captured in uncontrolled environments, the project explores data preparation techniques, model architecture, and training strategies to achieve high classification accuracy. Results demonstrate the effectiveness of handling class imbalance through both under-sampling and over-sampling methods, yielding robust models suitable for real-world applications. The under sampled model achieved a classification accuracy of 97.56% on a test set of 3,736 images. The over sampled model achieved far better results, achieving 99.786% accuracy on the same test set.