← Retour à l'épreuve

Kazakhstan FAIO 2025 · Énoncé

Kazakh Letters

Mis en forme par IOAI Training à partir de l'original (BeksultanSagyndyk/faio-tasks, licence Apache-2.0). L'original fait foi.

Overview

The Kazakh language uses a Cyrillic-based script with 42 distinct characters, including nine unique letters. While modern computer vision often relies on large, end-to-end deep learning models, this challenge places the spotlight on expert feature engineering and classical machine learning (ML) techniques.

Participants will not work with raw image files but with a curated, pre-processed tabular dataset where the visual information of each Kazakh letter image has been distilled into a set of handcrafted, numerical features.

Description

Your task is to build a classification model using a tabular dataset. This dataset contains a comprehensive set of features:

  • Statistical Features:
  • Mean intensity: Average pixel value.
  • Standard deviation: Pixel value spread.
  • Min/Max intensity: Helps capture extremes.
  • Skewness/Kurtosis: Measures asymmetry and peakedness of intensity distribution.
  • Histogram of pixel values: Divide intensity into bins (e.g., 16 bins) and count occurrences.
  • Shape-Based / Geometric Features:
    • Centroid: Center of mass of the digit.
    • Bounding box: Width, height, and aspect ratio.
    • Pixel density: Fraction of non-zero pixels.
    • Edge pixels count: Number of boundary pixels using edge detection.
    • Moments / Hu moments: 7 invariant moments describing shape, rotation, scale invariant.
    • Contours: Number of contours or contour areas using OpenCV.
    • Radial features: Distance of non-zero pixels from centroid.
    • Fourier descriptors of contours: Shape descriptors invariant to rotation/scale.
    • Symmetry measures: Compare left-right or top-bottom halves.
  • Projection / Profile Features:
  • Horizontal & vertical projections: Sum of pixel values along rows and columns.
  • Zoning features: Divide image into zones (e.g., 4x4) and calculate density in each zone.
  • Transform-Based Features
  • Fourier Transform: Frequency domain coefficients.
  • Discrete Cosine Transform (DCT): Common in compression, captures low-frequency patterns.
  • Wavelet Transform: Multi-scale decomposition, capturing edges and textures.

Files

  • train.csv - the training set. Contains the engineered features (e.g., f1, f2, ...) derived from the letter images, along with the true character label (target).
  • test.csv - the test set. Contains the same set of engineered features, but the target column is withheld for prediction.
  • sample_submission.csv - a sample submission file in the correct format.
  • labels.csv - supplemental information about the data.
kaggle competitions download -c qaz-letters