NoPainNoTrain - Artificial Neural Networks and Deep Learning Projects

Welcome to the repository for the NoPainNoTrain group projects in the Artificial Neural Networks and Deep Learning (AN2DL) course. This repository contains the code and reports for two group assignments where we achieved the highest possible score: 5.5/5.5 on both projects.

Team Members

• Noemi Bongiorni
• Alessandro Pedone
• Simone Licciardi
• Federico Maria Riva

---

Projects

1. Blood Cell Classification (Homework 1)

Objective:
Classify blood cells in RGB images into eight disjoint classes using a convolutional neural network (CNN).

Key Highlights:

• Dataset: 13,759 RGB images (96x96 resolution), filtered to 11,959 images after preprocessing.
• Methodology:
  • Built a baseline CNN architecture with regularization and augmentation.
  • Implemented Transfer Learning (TL) and Fine Tuning (FT) with architectures such as ConvNeXt and EfficientNet.
  • Deployed an ensemble of the best-performing models for improved accuracy and robustness.
• Final Results:
  • Codabench Score: 76%
  • Top Accuracy (ConvNeXtSmall): 96.20%

Find detailed information in the HW1 Report.

---

2. Mars Terrain Segmentation (Homework 2)

Objective:
Segment grayscale Mars terrain images into five classes: Background, Soil, Bedrock, Sand, and Big Rock.

Key Highlights:

• Dataset: 2,615 grayscale images (64x128 resolution), filtered to 2,505 images.
• Methodology:
  • Impemented Keras custom layer for Egde Decetions, Thesholding and others methods of computer vision
  • Explored a dual UNet architecture (Global and Local perspectives).
  • Designed a custom loss function combining Dice Loss, Focal Loss, and Boundary Loss.
  • Applied advanced data augmentation and optimization techniques.
• Final Results:
  • Mean IoU (Kaggle): 64.91% (Baseline: 32.81%)
• Kaggle Competition Link: AN2DL 2024-2025 Homework 2

Find detailed information in the HW2 Report.

---