Image augmentation transforms in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Image augmentation transforms

Problem:You have a small image dataset for a classification task. The current model trains well on the training images but performs poorly on new images, showing signs of overfitting.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Training loss: 0.15, Validation loss: 0.65

Issue:The model overfits the training data and does not generalize well to validation data due to limited and similar training images.

Your Task

Use image augmentation transforms to increase data variety and reduce overfitting, aiming to improve validation accuracy to at least 80% while keeping training accuracy below 92%.

Do not change the model architecture.

Only modify the data loading and preprocessing pipeline to add augmentation.

Use common augmentation techniques like rotation, flipping, and zoom.

Hint 1

Hint 2

Hint 3

Solution

Computer Vision

import tensorflow as tf
from tensorflow.keras import layers, models

# Load dataset
(train_images, train_labels), (val_images, val_labels) = tf.keras.datasets.cifar10.load_data()

# Normalize images
train_images = train_images / 255.0
val_images = val_images / 255.0

# Define image augmentation pipeline
data_augmentation = tf.keras.Sequential([
    layers.RandomFlip('horizontal'),
    layers.RandomRotation(0.1),
    layers.RandomZoom(0.1),
    layers.RandomTranslation(0.1, 0.1)
])

# Build model (same architecture as before)
model = models.Sequential([
    layers.Input(shape=(32, 32, 3)),
    data_augmentation,  # Apply augmentation only during training
    layers.Conv2D(32, (3,3), activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(64, (3,3), activation='relu'),
    layers.MaxPooling2D(),
    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(10, activation='softmax')
])

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Train model
history = model.fit(train_images, train_labels, epochs=20, batch_size=64, validation_data=(val_images, val_labels))

Added a data augmentation pipeline with random horizontal flip, rotation, zoom, and translation.

Inserted the augmentation layer at the start of the model to augment training images on the fly.

Kept validation data unchanged to properly measure generalization.

Results Interpretation

Before augmentation: Training accuracy was 95%, validation accuracy was 70%. The model overfitted the training data.

After augmentation: Training accuracy decreased to 90%, validation accuracy improved to 82%. Loss values also show better generalization.

Image augmentation artificially increases data variety, helping the model learn more robust features and reducing overfitting. This improves validation accuracy by making the model better at handling new, unseen images.

Bonus Experiment

Try adding color jitter (random brightness and contrast changes) to the augmentation pipeline and observe if validation accuracy improves further.

💡 Hint

Use layers.RandomBrightness and layers.RandomContrast from TensorFlow or implement custom augmentation functions.

Practice

1. What is the main purpose of image augmentation in training machine learning models?

easy

A. To reduce the size of the training dataset

B. To remove noise from images

C. To create more varied training images by modifying originals

D. To convert images to grayscale only

Solution

1. Step 1: Understand image augmentation

   Image augmentation means making small changes to original images to create new ones.

2. Step 2: Purpose in training

   This helps models see more variety and learn better, avoiding overfitting.

3. Final Answer:

   To create more varied training images by modifying originals -> Option C

4. Quick Check:

   Image augmentation = create varied images [OK]

Hint: Augmentation means changing images to get more training data [OK]

Common Mistakes:

• Thinking augmentation reduces dataset size
• Confusing augmentation with noise removal
• Assuming augmentation only changes color

2. Which of the following is the correct way to apply a horizontal flip using PyTorch's torchvision transforms?

easy

A. transforms.RandomHorizontalFlip(p=1.0)

B. transforms.HorizontalFlip()

C. transforms.FlipHorizontal()

D. transforms.RandomFlip(direction='horizontal')

Solution

1. Step 1: Recall torchvision syntax

   PyTorch uses transforms.RandomHorizontalFlip(p=probability) to flip images horizontally.

2. Step 2: Check options

   Only transforms.RandomHorizontalFlip(p=1.0) matches the correct function and parameter style.

3. Final Answer:

   transforms.RandomHorizontalFlip(p=1.0) -> Option A

4. Quick Check:

   Correct PyTorch flip = RandomHorizontalFlip [OK]

Hint: Look for 'RandomHorizontalFlip' with probability parameter [OK]

Common Mistakes:

• Using non-existent transform names
• Missing the probability parameter
• Confusing horizontal with vertical flip

3. Given the following code snippet using torchvision transforms, what is the output image size after applying the transforms?

transform = transforms.Compose([
    transforms.Resize((128, 128)),
    transforms.RandomCrop(100),
    transforms.ToTensor()
])

image = Image.open('sample.jpg')
output = transform(image)
print(output.shape)

medium

A. [3, 128, 128]

B. [3, 100, 100]

C. [1, 100, 100]

D. [3, 228, 228]

Solution

1. Step 1: Analyze each transform step

   First, image is resized to 128x128 pixels with 3 color channels (RGB). Then a random crop of size 100x100 is taken.

2. Step 2: Determine output tensor shape

   After cropping, the image size is 100x100 with 3 channels. ToTensor() converts it to a tensor with shape [channels, height, width] = [3, 100, 100].

3. Final Answer:

   [3, 100, 100] -> Option B

4. Quick Check:

   Resize then crop = final size 100x100 [OK]

Hint: Resize then crop means output size = crop size [OK]

Common Mistakes:

• Ignoring the crop step size
• Confusing channel dimension with batch size
• Assuming crop keeps original size

4. The following code is intended to rotate an image by 45 degrees using torchvision transforms, but it raises an error. What is the mistake?

transform = transforms.Compose([
    transforms.Rotate(45),
    transforms.ToTensor()
])

image = Image.open('sample.jpg')
output = transform(image)

medium

A. transforms.Rotate doesn't exist; should use transforms.functional.rotate or transforms.RandomRotation

B. The angle 45 must be in radians, not degrees

C. ToTensor must come before Rotate

D. Image.open returns a tensor, so transform fails

Solution

1. Step 1: Check torchvision transform names

   There is no transforms.Rotate class. Rotation is done with transforms.RandomRotation or using functional API.

2. Step 2: Identify correct usage

   To rotate by a fixed angle, use transforms.RandomRotation([45, 45]) or transforms.functional.rotate. The code as is will cause an AttributeError.

3. Final Answer:

   transforms.Rotate doesn't exist; should use transforms.functional.rotate or transforms.RandomRotation -> Option A

4. Quick Check:

   No transforms.Rotate in torchvision [OK]

Hint: Check transform names carefully; Rotate is not a direct class [OK]

Common Mistakes:

• Using non-existent transform classes
• Confusing degrees and radians
• Wrong order of transforms

5. You want to augment a dataset of images to improve model robustness. Which combination of transforms would best simulate real-world variations while keeping image size constant?

hard

A. transforms.RandomCrop(224), transforms.RandomRotation(180), transforms.Resize(128)

B. transforms.Resize(256), transforms.CenterCrop(224), transforms.RandomVerticalFlip() only

C. transforms.RandomRotation(90), transforms.RandomCrop(200), transforms.ToTensor()

D. transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ColorJitter(brightness=0.2, contrast=0.2)

Solution

1. Step 1: Understand augmentation goals

   We want to simulate real-world changes like size, flip, and color while keeping output size fixed.

2. Step 2: Evaluate options

   transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ColorJitter(brightness=0.2, contrast=0.2) resizes and crops randomly to 224x224, flips horizontally, and changes brightness/contrast, all common augmentations that keep size constant.

3. Step 3: Check other options

   transforms.Resize(256), transforms.CenterCrop(224), transforms.RandomVerticalFlip() only flips vertically and crops but lacks color changes. transforms.RandomRotation(90), transforms.RandomCrop(200), transforms.ToTensor() changes size unpredictably and transforms.RandomCrop(224), transforms.RandomRotation(180), transforms.Resize(128) resizes after cropping, changing size.

4. Final Answer:

   transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ColorJitter(brightness=0.2, contrast=0.2) -> Option D

5. Quick Check:

   Best augmentations keep size fixed and add variety [OK]

Hint: Pick transforms that keep size fixed and add flip + color changes [OK]

Common Mistakes:

• Choosing transforms that change image size unpredictably
• Ignoring color augmentations
• Using only vertical flips which are less common