Artificial Neural Networks and Deep Learning 2023 - Image Classification Challenge

This repository contains materials for the first Homework of the course Artificial Neural Networks and Deep Learning at Politecnico di Milano. The assignment focuses on binary image classification, specifically classifying plants into two categories based on their health state: "healthy" and "unhealthy." The task involves leveraging deep learning techniques to predict the correct class label for each image in the provided dataset.

Dataset Details:

• Image size: 96x96
• Color space: RGB
• File Format: npz
• Binary Classes:
  • 0: "healthy"
  • 1: "unhealthy"
• Data: A numpy array of shape 5200x96x96x3, containing the RGB images.
• Labels: A 1-dimensional numpy array of shape 5200 with values in {'healthy', ‘unhealthy'}.

Models Details:

The classification task employs a Transfer Learning approach with the ConvNeXtLarge model. The initialization is made with pre-trained weights followed by fine-tuning in two phases. To face the moderately-sized dataset some data augmentation techniques are applied, including:

• RandomFlip
• Crop
• ZeroPadding
• Brightness
• Contrast

For a detailed understanding of the model architecture and techniques employed, please refer to the provided Report.

Results:

With the proposed model we obtained 4.5/5 in the challenge score.
The model have been tested on an hidden dataset and it reached a 78% of accuracy.