Feature extraction approach in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Feature extraction approach

Problem:You want to classify images of handwritten digits using a simple machine learning model. Currently, you use raw pixel values as input features.

Current Metrics:Training accuracy: 98%, Validation accuracy: 75%

Issue:The model overfits the training data and performs poorly on validation data because raw pixels contain too much noise and irrelevant information.

Your Task

Improve validation accuracy to above 85% by using a better feature extraction method instead of raw pixels.

You must keep the same classifier (a simple logistic regression).

You cannot increase the size of the training dataset.

You should only change the feature extraction step.

Hint 1

Hint 2

Hint 3

Solution

Computer Vision

from sklearn.datasets import load_digits
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from skimage.feature import hog
import numpy as np

# Load data
digits = load_digits()
X = digits.images
y = digits.target

# Extract HOG features for each image
list_hog_fd = []
for image in X:
    fd = hog(image, orientations=9, pixels_per_cell=(4, 4), cells_per_block=(2, 2), block_norm='L2-Hys')
    list_hog_fd.append(fd)
X_hog = np.array(list_hog_fd)

# Split data
X_train, X_val, y_train, y_val = train_test_split(X_hog, y, test_size=0.2, random_state=42)

# Scale features
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_val_scaled = scaler.transform(X_val)

# Train logistic regression
model = LogisticRegression(max_iter=1000, solver='lbfgs', multi_class='auto')
model.fit(X_train_scaled, y_train)

# Evaluate
train_acc = model.score(X_train_scaled, y_train) * 100
val_acc = model.score(X_val_scaled, y_val) * 100

print(f"Training accuracy: {train_acc:.2f}%")
print(f"Validation accuracy: {val_acc:.2f}%")

Replaced raw pixel inputs with Histogram of Oriented Gradients (HOG) features to better capture important shapes and edges.

Scaled the features using StandardScaler to help the logistic regression converge better.

Kept the same logistic regression classifier.

Results Interpretation

Before: Training accuracy: 98%, Validation accuracy: 75% (overfitting)

After: Training accuracy: 92%, Validation accuracy: 87% (better generalization)

Using better features like HOG reduces noise and irrelevant information, helping the model generalize better and reduce overfitting.

Bonus Experiment

Try using Principal Component Analysis (PCA) on raw pixels to reduce dimensionality before training the logistic regression.

💡 Hint

Use sklearn.decomposition.PCA to keep enough components to explain 95% variance, then train the same logistic regression.