Data Augmentation

Data augmentation refers to techniques that artificially expand a training dataset by creating modified versions of existing samples. These methods help machine learning models generalize better, reduce overfitting, and perform well on new, unseen data. The process mimics real-world variations that might occur in the data, making models more resilient to different conditions and scenarios.

Core Concepts and Implementation

Image Augmentation Techniques

Geometric Transformations:

# Common geometric transforms
def geometric_augment(image):
    # Rotation: -45 to +45 degrees
    rotated = rotate(image, angle=random.uniform(-45, 45))

    # Scale: 0.8x to 1.2x
    scaled = resize(image, scale=random.uniform(0.8, 1.2))

    # Flip: horizontal/vertical
    flipped = np.fliplr(image)  # horizontal flip

    # Shear: -20 to +20 degrees
    sheared = shear(image, angle=random.uniform(-20, 20))

    return [rotated, scaled, flipped, sheared]

Color Space Modifications:

def color_augment(image):
    # Brightness: ±30%
    brightness = adjust_brightness(image, factor=random.uniform(0.7, 1.3))

    # Contrast: ±30%
    contrast = adjust_contrast(image, factor=random.uniform(0.7, 1.3))

    # Saturation: ±30%
    saturation = adjust_saturation(image, factor=random.uniform(0.7, 1.3))

    # Hue: ±15 degrees
    hue = adjust_hue(image, factor=random.uniform(-15, 15))

    return [brightness, contrast, saturation, hue]

Advanced Image Augmentations:

1. Mixing Methods:

   def mixup(image1, image2, label1, label2, alpha=0.2):
       # Create weighted combinations
       lambda_val = np.random.beta(alpha, alpha)
       mixed_image = lambda_val * image1 + (1 - lambda_val) * image2
       mixed_label = lambda_val * label1 + (1 - lambda_val) * label2
       return mixed_image, mixed_label

2. Cutout/Random Erasing:

   def cutout(image, size=50):
       # Create random masks
       h, w = image.shape[:2]
       x = np.random.randint(w - size)
       y = np.random.randint(h - size)
       image[y:y+size, x:x+size] = 0
       return image

Text Augmentation Techniques

Basic Text Operations:

def text_augment(text):
    augmentations = []

    # Synonym replacement
    synonyms = wordnet.synsets(word)
    new_text = replace_word(text, word, synonyms.lemmas()[0].name())
    augmentations.append(new_text)

    # Back translation
    intermediate = translate(text, 'en', 'fr')
    back_translated = translate(intermediate, 'fr', 'en')
    augmentations.append(back_translated)

    # Word insertion/deletion
    words = text.split()
    inserted = words.copy()
    inserted.insert(random.randint(0, len(words)), random_word)
    augmentations.append(' '.join(inserted))

    return augmentations

Advanced Text Methods:

# Contextual augmentation using language models
def contextual_augment(text, model):
    # Mask random words
    masked_text = mask_random_words(text)

    # Predict replacements
    predictions = model(masked_text)

    # Sample from top-k predictions
    augmented = replace_masks(masked_text, predictions, k=5)
    return augmented

Audio Augmentation

Time-Domain Transformations:

def audio_augment(waveform, sr=16000):
    augmentations = []

    # Time stretching
    stretched = librosa.effects.time_stretch(waveform, rate=1.2)

    # Pitch shifting
    pitched = librosa.effects.pitch_shift(waveform, sr=sr, n_steps=2)

    # Adding noise
    noise = np.random.normal(0, 0.005, len(waveform))
    noisy = waveform + noise

    augmentations.extend([stretched, pitched, noisy])
    return augmentations

Application-Specific Examples

Medical Imaging

def medical_image_augment(scan):
    # Preserve diagnostic features
    augmentations = []

    # Intensity windowing
    windowed = apply_window_level(scan, window=2000, level=0)

    # Elastic deformation
    deformed = elastic_transform(scan, alpha=100, sigma=10)

    # Noise addition (within diagnostic limits)
    noise_level = calculate_safe_noise_level(scan)
    noisy = add_controlled_noise(scan, noise_level)

    augmentations.extend([windowed, deformed, noisy])
    return augmentations

Satellite Imagery

def satellite_augment(image):
    # Weather and atmospheric effects
    augmentations = []

    # Cloud coverage simulation
    cloudy = add_cloud_overlay(image, coverage=0.3)

    # Shadow simulation
    shadowed = add_shadow_mask(image, angle=45)

    # Time of day simulation
    time_shifted = adjust_lighting(image, time_of_day=14)

    augmentations.extend([cloudy, shadowed, time_shifted])
    return augmentations

Quality Control and Validation

Augmentation Verification:

def verify_augmentation(original, augmented, threshold=0.7):
    # Check structural similarity
    ssim_score = structural_similarity(original, augmented)

    # Verify label preservation
    label_preserved = verify_labels(original, augmented)

    # Check feature consistency
    feature_similarity = compare_features(original, augmented)

    return all([
        ssim_score > threshold,
        label_preserved,
        feature_similarity > threshold
    ])

Performance Impact Metrics

Common measurements for augmentation effectiveness:

def evaluate_augmentation(model, original_data, augmented_data):
    metrics = {
        'accuracy_improvement': 0,
        'generalization_score': 0,
        'overfitting_reduction': 0
    }

    # Train with and without augmentation
    baseline = train_model(model, original_data)
    augmented = train_model(model, augmented_data)

    # Compare performance
    metrics['accuracy_improvement'] = augmented.accuracy - baseline.accuracy
    metrics['generalization_score'] = compare_validation_curves(baseline, augmented)
    metrics['overfitting_reduction'] = compare_overfitting_metrics(baseline, augmented)

    return metrics

These techniques continue to evolve as researchers develop new methods to create more effective and domain-specific data augmentations.