Research

Deep Learning Based Custom Image Multi-Class Classifier

M.Sc. Artificial Intelligence and Machine Vision report, University of East London

Summary

A CIFAR-10 classifier built on VGG19 with transfer learning and data augmentation, reaching 91.3% test accuracy.

Abstract

This study presents the development and evaluation of a custom multi-class image classifier on the CIFAR-10 dataset: 60,000 32×32 colour images in ten categories. It investigates how architectural decisions and hyperparameters (layers, filter sizes, dropout, batch size and learning rate) affect performance, building on the VGG19 architecture with transfer learning, and runs experiments on data augmentation, learning-rate schedules and epochs. Deeper architectures combined with proper augmentation significantly improved accuracy; the final model reached 91.3% on the test set, with the confusion matrix pointing to visually similar classes as the main remaining errors.