Supervised Project Report

Phase 1

Dataset preparation

The German Traffic Sign Recognition Benchmark (GTSRB) dataset was chosen.

What was done

1. Dataframe construction
   • for every sample, translated its image file into a 3D array (with multiple channels)

---

EDA

Exploratory Data Analysis was conducted to understand the distribution and variance within the visual data.

What was done

1. Class Balance Analysis
   • Plotted the number of samples per traffic sign class (0 to 42) using a bar chart to observe potential class imbalances.
2. Raw Image Visualization & Region of Interest (ROI)
   • Extracted the bounding box coordinates (Roi.X1, Roi.Y1, Roi.X2, Roi.Y2) from the dataset to identify the specific region of interest in each image.
3. Lighting Condition Analysis
   • Visualized the distribution of image brightness across classes using a boxplot, showing heavy variances in contrast and lighting.
     

---

Features Preparation

To prepare the dataset for both traditional and deep learning classifiers, it needed some work

What was done

1. Cropping and Resizing (for deep learning)
   • Cropped each image to its exact ROI bounding box.
     • to remove background noise.
   • Standardized all cropped images to a fixed $32\times 32$ resolution.
     
2. Flattening
   • Flattened the images into a 1D array.
3. Scaling (for both)
   • Scaled the pixel values using a normalizer.

---

PCA

Implemented PCA and picked the one to use, using its explained variance plot.

What was done

1. Initial PCA transformer
   • Initially chose 50 as the # of principal components to keep, but…
2. Choosing the one
   • its variance plot showed that 30 components were sufficient (they explained 90%).
   • thus, I switched to one with 30 components.

---

LDA

Implemented LDA as an alternative for class-aware dimensionality reduction.

What was done

1. Maximizing Class Separation
   • Applied LDA with 10 components.
   • Plotted the cumulative variance to visualize how well the discriminant indices separated the traffic sign classes.

NOTE

• from this point on, the PCA transformed features were used in ML classifiers.
• the # of samples also was reduced to 5000, since ML classifiers don’t use the GPU and take way TOO long to run.

---

KNN

A K-Nearest Neighbors classifier was implemented from scratch.

What was done

1. Grid Search Optimization
   • Ran GridSearchCV testing $k=[1,3,5,7]$.
   • Discovered that $k=1$ yielded the best parameters.

Performance

• Accuracy
  • 76.50%
• F1
  • 0.7636
• ROC AUC
  • 0.8781

Not the worst performance, especially since only a subset of the data was used.

---

Gaussian Naive Bayes

A Gaussian Naive Bayes classifier was implemented from scratch.

What was done

1. Training (Mean & Variance)
   • Grouped the training data by class and calculated the mean ($\mu$) and variance (${\sigma }^2$) for every feature.
   • Added a small epsilon value (1e-9) to the variance to prevent numerical instability.
2. Predicting (Log Posteriors)
   • used log-likelihoods.

Performance

• Accuracy
  • 18.33%
• F1
  • 0.1965
• ROC AUC
  • 0.6805

Its performance was really terrible. why?

• gaussian naive bayes assumes that all the features (pixel values) are independent.
• pixels within the same neighborhood are highly dependent.

---

Ensemble

To improve the stability of the scratch models, an ensemble approach was applied.

What was done

1. Bagging
   • Wrapped the custom $k=1$ KNN model inside a BaggingClassifier.
   • Trained 5 estimators on random subsets of the data (bootstrapping) to reduce variance and output an aggregated prediction.

---

Phase 2

CNN

A Convolutional Neural Network (CNN) was implemented from scratch to natively handle the 2D spatial structures within the images.

What was done

1. Architecture Design

   • Normalization
     • Initially normalized the training pixels with a Normalization layer.
   • Level 1 block
     • Used $3\times 3$ Conv2D layers (32 filters).
     • this extracts the lowest level features.
   • Level 2 block
     • Deeper $3\times 3$ Conv2D layers (64 filters).
     • this extracts higher-level traffic sign symbols.
   • Classification Head
     • Flattened the maps into a 512-neuron Dense layer with severe Dropout(0.5) to prevent overfitting, outputting to a 43-class softmax.
   • used BatchNormalization to efficiently normalize features after every block.

   

Performance

• Accuracy
  • 97.85%
• Loss
  • 0.0898

---

Optimizer Comparison

Optimized the learning rate decay during the CNN’s training phase.

What was done

1. Learning Rate Scheduling
   • Implemented ReduceLROnPlateau.
   • Set to monitor val_loss, reducing the learning rate by a factor of 0.2 if the loss plateaus for 3 epochs (with a floor of 1e-6).

---

Transfer Learning

Benchmarked the scratch CNN against heavy, pre-trained architectures.

What was done

1. Pretrained vs Fine-Tuned
   • Loaded ResNet50 and MobileNetV2 with ImageNet weights.
   • Tested both in a purely feature-extraction state (pretrained) and a partially unfrozen state (fine-tuned).
2. Performance Conclusion
   • Both transfer learning architectures underperformed compared to the custom CNN.
   • Hypothesis
     • The pretrained models were optimized for large, high-resolution images, whereas the custom CNN was strictly constrained and tailored for the $32\times 32$ matrices of the GTSRB dataset.

---

AutoEncoders

An unsupervised Autoencoder pipeline was built to learn visual representations, denoise inputs, and flag anomalous data.

What was done

1. Image Enhancement
   • Preprocessed the images using CLAHE (Contrast Limited Adaptive Histogram Equalization) to improve local contrast before feeding them to the network.
2. Denoising Architecture
   • Injected artificial Gaussian noise (noise_factor = 0.1) into the training images.
   • Built a bottleneck architecture with Convolutional and MaxPooling downsampling (Encoder) and UpSampling2D (Decoder).
   • Trained the model mapping the noisy input directly to the clean target, optimizing via Mean Absolute Error (MAE).
3. Anomaly Detection
   • Reconstructed the test set and calculated the Mean Squared Error (MSE) of the reconstructions.
   • Established a 95th-percentile threshold. Images exceeding this reconstruction error threshold were flagged as “Anomalies.”
   • Visualized the “Best” (normal) vs “Worst” (anomalous) reconstructions to verify the system’s filtering logic.