SVHN Digit Classification with MobileNetV2 Transfer Learning

SVHN Digit Classification

Leveraging Transfer Learning with MobileNetV2

Interactive Agenda

1. Dataset & Exploration

2. Preprocessing & Model

3. Training Analysis

4. Optimization & Deploy

Introduction & Task Overview

The Goal: Classify digits (0-9) from the SVHN Cropped dataset using Transfer Learning.

• Dataset: SVHN (Street View House Numbers) • Challenge: Real-world clutter, variable lighting • Approach: MobileNetV2 Pretrained Backbone • Outcome: Evaluated & Deployed Model

Dataset Exploration (EDA)

73,257

Training Examples

26,032

Test Examples

Preprocessing Pipeline

IMG_SIZE = 96 def format_image(image, label): # Resize to match MobileNetV2 input image = tf.image.resize(image, (IMG_SIZE, IMG_SIZE)) # Normalize (0-255 -> 0.0-1.0) image = image / 255.0 return image, label # Pipeline optimization train = raw_train.map(format_image).cache().shuffle(1000).batch(32)

Key Operations:

• Resizing: 96x96 pixels • Normalization: Scaling pixel intensity • Augmentation: Random Rotation & Zoom • Caching: Optimized data throughput

Model Architecture

We employ a Transfer Learning strategy using MobileNetV2 features with a custom classification head.

Implementation (Keras)

# Base Model (Pretrained) base_model = tf.keras.applications.MobileNetV2( input_shape=(IMG_SIZE, IMG_SIZE, 3), include_top=False, weights='imagenet' ) base_model.trainable = False # Freeze weights # Custom Head model = tf.keras.Sequential([ base_model, tf.keras.layers.GlobalAveragePooling2D(), tf.keras.layers.SeparableConv2D(1024, (3,3), padding='same'), tf.keras.layers.Dropout(0.2), tf.keras.layers.Dense(10, activation='softmax') ])

Note: Feature extraction frozen to leverage ImageNet features.

Training Configuration

• Optimizer: Adam (LR=1e-3) • Loss: Sparse Categorical Crossentropy • Epochs: 3 • Steps per Epoch: 50 (Constraint)

• EarlyStopping (Patience=2) • ReduceLROnPlateau (Factor=0.3) • Validation Split: 10%

Baseline Results (3 Epochs)

Model shows steady convergence even within limited training steps. Accuracy climbed from 38% to 46%.

Model Evaluation

49.82%

Test Set Accuracy

Confusions Noted: • Digits 2 and 3 usually confused • Varied lighting impacts confidence

Optimization: Learning Rate Tuning

Reducing the Learning Rate to 1e-4 significantly improved convergence, yielding >51% accuracy by Epoch 2.

Deployment & Prediction

def predict_image(img_path): # Save and Load model model.save('svhn_mobilenetv2.keras') # Preprocess single image img = load_and_resize(img_path) img_batch = tf.expand_dims(img, 0) # Inference preds = model.predict(img_batch) score = tf.nn.softmax(preds[0]) return np.argmax(score), 100 * np.max(score)

Ethical AI Considerations

Bias & Generalization

Model may fail on house numbers with unusual fonts or backgrounds not present in SVHN training data.

Compute Efficiency

Using Transfer Learning (MobileNetV2) reduces energy consumption compared to training from scratch.

Conclusion & Future Work

Successfully implemented a transfer learning pipeline achieving >50% accuracy on a challenging 10-class dataset with minimal training.

Q & A