Multi-class classification using HOG + LBP feature fusion with SVM

Classify brain MRI scans into one of four categories using classical machine learning (without any deep learning or pre-trained models)

Brain Tumor — Kaggle Hackathon Link

Imbalanced dataset - tumor classes make up ~70% of training data, with varying distribution across glioma, meningioma, and pituitary types. Labels are embedded directly in filenames (e.g. 00001_image(45)_glioma_tumor.png).

Labels were parsed directly from filenames by matching known class suffixes. No separate label file was required.

00001_gg(399)_glioma_tumor.png  →  glioma_tumor

• All images resized to 128×128 pixels
• Converted to grayscale
• Normalized to [0, 1] for consistent feature extraction

Two complementary feature descriptors were combined into a single 8,260-dimensional feature vector.

HOG (Histogram of Oriented Gradients) -> captures edge and shape structure

• 9 orientations
• 8×8 pixels per cell
• 2×2 cells per block
• L2-Hys normalization

LBP (Local Binary Pattern) - captures local texture patterns

• P=8, R=1, uniform method
• Applied over a 4×4 spatial grid

HOG and LBP are complementary -> HOG captures global shape/edge structure while LBP captures fine-grained local texture. Their fusion produces a richer feature representation for MRI classification.

Combined feature vector: HOG + LBP = 8,260 dims

Horizontal flips were applied to training images only (not validation), doubling the effective training size.

• StandardScaler applied before PCA (required for SVM)
• Top 300 principal components retained
• Explained variance retained: 69.54%

Reduces the 8,260-dim feature space significantly while preserving the most discriminative structure, speeding up SVM training without sacrificing meaningful signal.

• class_weight="balanced" applied to SVM -> penalizes minority class misclassifications proportionally
• Stratified 80/20 train/validation split -> ensures equal class representation across both sets

SVM with RBF kernel tuned via 5-fold stratified cross-validation.

Search space:

Best parameters found:

Final model: SVM · RBF kernel · C=10 · gamma=scale

Evaluated on validation set - 402 original held-out images

Brain MRI Tumor Classification
├── NOTICE
├── LICENSE 
├── README.md 
├── report.txt                              -> Whole Approach Report
├── submission.csv                          -> 863-row prediction file with `image_id` and `label` columns
└── brain-mri-tumor-classification.ipynb    -> Full notebook - preprocessing, feature extraction, training, and evaluation