# SVHN Digit Classification with MobileNetV2 Transfer Learning
> Learn how to classify Street View House Numbers (SVHN) using transfer learning with MobileNetV2, Keras, and TensorFlow. Includes preprocessing and code samples.

Tags: transfer-learning, mobilenetv2, computer-vision, deep-learning, tensorflow, keras, svhn-dataset, image-classification
## SVHN Digit Classification with MobileNetV2
Presentation on leveraging transfer learning for digit recognition in real-world images.

## Dataset Exploration (EDA)
* **Dataset:** SVHN (Street View House Numbers).
* **Training Examples:** 73,257.
* **Test Examples:** 26,032.
* **Challenge:** Real-world clutter and variable lighting conditions.

## Preprocessing Pipeline
* **Image Size:** Resized to 96x96 pixels.
* **Normalization:** Scaling pixel intensities to [0.0, 1.0].
* **Operations:** Random augmentation (rotation/zoom), caching, and shuffling for throughput optimization.

## Model Architecture & Implementation
* **Backbone:** Pretrained MobileNetV2 (weights from ImageNet).
* **Strategy:** Feature extraction with frozen weights (`trainable = False`).
* **Custom Head:** GlobalAveragePooling2D, SeparableConv2D (1024 filters), Dropout (0.2), and a Dense output layer with Softmax activation for 10 classes.

## Training Configuration
* **Optimizer:** Adam (Initial Learning Rate = 1e-3).
* **Loss Function:** Sparse Categorical Crossentropy.
* **Callbacks:** EarlyStopping (Patience=2) and ReduceLROnPlateau.

## Results and Performance
* **Baseline Accuracy:** Reached ~46% in 3 epochs.
* **Optimized Results:** Tuning the learning rate to 1e-4 improved convergence to >51% accuracy.
* **Test Accuracy:** Final test set evaluation at 49.82%.

## Ethical Considerations
* **Bias:** Potential failures on house numbers with unusual fonts.
* **Efficiency:** Transfer learning significantly reduces compute energy consumption compared to training from scratch.
---
This presentation was created with [Bobr AI](https://bobr.ai) — an AI presentation generator.