Overview

• Dataset Creation & Preprocessing

  • Gather at least 500 images for each of 5 or more distinct classes.

  • Assign correct class labels and maintain a clear folder structure.

  • Remove duplicates, low-quality images, and irrelevant content.

  • (You don’t need to split the dataset, as you can do this in the code.)

• Model from Scratch

  • Design a CNN with at least three convolutional blocks (Conv→Activation→Pooling).

  • Include regularisation (batch normalisation or dropout).

  • Train for a minimum of 20 epochs using an appropriate optimiser and learning-rate schedule.

  • Plot training and validation loss/accuracy curves.

• Transfer Learning

  • Select a pretrained model (e.g., MobileNet or ResNet).

  • Replace its classification head to match your dataset’s number of classes.

  • Freeze the backbone and train only the new head.

  • Unfreeze some or all backbone layers and fine-tune.

  • Use online data augmentation to improve generalisation.

  • Train for at least 10 epochs and record learning curves for each configuration, namely:

    • feature extraction (only the new head) with and without augmentation,

    • fine-tuning (unfreeze some or all layers) with and without augmentation.

• Analysis & Reporting

  • Tabulate final metrics: confusion matrix, accuracy, precision, and recall.

  • Compare convergence speed (epochs to plateau), final performance, and computational cost.

  • Discuss the advantages and limitations of each method in this context.