This is the repository for the Homeworks of Artificial Neural Networks and Deep Learning in the academic year 2024/2025 at Polytechnic of Milan.

Subject: 054307 - Artificial Neural Networks And Deep Learning

Professors: Boracchi Giacomo and Matteucci Matteo

Academic Year: 2024/2025

The homeworks includes two different projects involving neural networks and two different tasks:

• First Homework: Image Classification
• Second Homework: Semantic Segmentation

In this assignment, you will classify 96x96 RGB images of blood cells. These images are categorized into eight classes, each representing a particular cell state. This is a multi-class classification problem, so your goal is to assign the correct class label to each RGB image.

Blood Cells

To enlarge the dataset and to make the proposed neural network robust to changes in the dataset, the technique called Augmentation was used.

Augmentation 1

Augmentation 2

The implemented network is a ConvNeXtBase-based transfer learning model for multi-class image classification, featuring a pretrained backbone with custom fully connected layers to enhance feature representation and regularization.

ConvNeXtBase:

• Pretrained on ImageNet, used as a feature extractor.
• include_top=False excludes the original classification head.
• Frozen weights to preserve learned low- and mid-level features.

Global Average Pooling:

• Reduces spatial dimensions while retaining global feature information. Batch Normalization:
• Stabilizes training and accelerates convergence. Dropout Layers:
• Two stages of dropout (0.2 and 0.25) reduce overfitting. Fully Connected (Dense) Layers:
• Four layers of decreasing size: 512 → 256 → 128 → 64 neurons.
• ReLU activation introduces non-linearity.
• Optional BatchNormalization (commented out) for further stabilization.

Dense Layer with Softmax:

• Produces class probabilities for NUM_CLASSES.

Loss Function:

• Categorical cross-entropy with label smoothing (0.1) to improve generalization. Optimizer:
• AdamW with Cosine Decay Restarts for adaptive learning rate scheduling. Metrics:
• Accuracy, Precision, and Recall for comprehensive evaluation.

• Pretrained ConvNeXtBase backbone: leverages rich, hierarchical features from ImageNet.
• Custom top layers: provide flexibility for different classification tasks.
• Dropout and BatchNorm: reduce overfitting and improve training stability.
• Mixed precision training: speeds up computation and reduces memory usage.

This combination makes the architecture robust, efficient, and suitable for high-performance image classification tasks.
It was first trained on the dataset as is and next transfer leanring with fine tuning was performed by unfreezing the layers of the network.

A validation accuracy of 85.7% has been obtained after fine tuning. The technique called Test-Time Augmetation was also used to improve the performances.
For further details, check the report.
Some results of the trained networks for comparison:

In this assignment, you will receive 64x128 grayscale real images from Mars terrain. Pixels in these images are categorized into five classes, each representing a particular type of terrain. This is a semantic segmentation problem, so your goal is to assign the correct class label to each mask pixel.
Pretrained models are forbidden.

Mars Terrain

Mask

We used Albumentations augmentation pipeline to augment the dataset creating a dataset which is the union of smaller datasets created trhough augmentation to increase variability and the number of samples.

The implemented network is an enhanced U-Net for image classification/segmentation, integrating several advanced deep learning modules to improve feature extraction, multi-scale context understanding, and attention.

• Conv Blocks & ResNet Blocks:
  • Four stages of convolutional processing with residual connections.
  • Each block increases feature depth (32 → 64 → 128 → 256).
  • Residual shortcuts stabilize training and allow deeper representation learning.
• Squeeze-and-Excitation (SE) Block:
  • Recalibrates channel-wise feature responses by modeling interdependencies between channels.
  • Helps the network focus on the most informative features.

• Multi-Scale ASPP:
  • Atrous Spatial Pyramid Pooling (ASPP) applied at multiple scales (filters, filters/2, filters/4).
  • Captures both local fine details and global context by combining dilated convolutions with different rates.
• Concatenation with skip features:
  • Deeper features are fused with intermediate representations (e.g., from d3 and d4) for stronger context.
• 3×3 Conv Layer:
  • Further processing with 512 filters to unify features before decoding.

• Attention Gates:
  • Applied at each skip connection to highlight relevant features and suppress irrelevant background noise.
  • Gating signals come from ASPP-enhanced encoder outputs.
• Conv2DTranspose Layers (Deconvolutions):
  • Used for upsampling at each stage (256 → 128 → 64 → 32).
• Weighted Fusion:
  • Combines skip connections and upsampled features adaptively with learned importance weights.
• ResNet Blocks:
  • Each upsampling stage includes a residual block to refine features.

• Fusion Block:
  • Upsamples and unifies multi-scale outputs (u1, u2, u3, u4) into a single feature representation.
  • Uses weighted union to balance contributions from different scales.
• Output Layer:
  • Final 1×1 Conv2D layer with Softmax activation for multi-class classification (5 classes).

• ResNet Blocks: enable deep feature learning with stable gradients.
• Squeeze-and-Excitation: channel attention mechanism to boost 805B informative filters.
• ASPP & Multi-Scale ASPP: powerful context aggregation at multiple dilation rates.
• Attention Gates: dynamic feature selection at skip connections.
• Fusion Block: combines multi-scale decoder outputs into a unified representation.

This combination makes the architecture robust, context-aware, and highly discriminative, ideal for structured visual tasks.

We used the mean IoU (Intersection over Union) as the metric to evaluate the results.
Categorical Crossentropy is the loss function. Weights are used to weight differently the classes, to take into account under-represented classes.

Mask 1

Mask 2

The final validation mean_IoU is 72.95%.
We used also Test-Time Augmentation to improve the final result.
For further details, check the report.
Some results for different loss functions:

First Homework Final Position:

• Private Leaderboard (Final phase): Top 20 over 700+ people

Second Homework Final Position:

• Public Leaderboard (Intermediate position): 63/197 (197 groups in total)
• Private Leaderboard (Final Position): 9/197

This Project was developed by:

• Corradina Davide @CorraPiano
• De Introna Federico @federicodeintrona
• Di Giore Francesco @Digioref