thekamilya/kazakh-printed-dataset

Kazakh Printed Dataset for OCR task

Data Lineage

This dataset was synthetically generated using issai/kazparc as the base. Since Kazakh OCR data is scarce, I developed a pipeline to transform digital Kazakh text into a printed-style dataset.

Generation Process

Source:

Text samples were extracted from issai/kazparc.

Augmentation & Stylization:

• —Random Background Color: Simulates different lighting conditions by alternating between high-contrast "light mode" (black on white) and "dark mode" (white on black) color schemes.
• —Dynamic Image Size: Automatically calculates the canvas width based on the length of the input text to ensure a realistic aspect ratio.
• —Random Font & Size: Injects stylistic variety by selecting random typeface files and varying the font size between 40 and 60 points.
• —Center Text with Random Jitter: Places the text in the center of the image but applies small, random horizontal and vertical offsets to mimic human inconsistency.
• —Augmentations (Rotation): Tilts the entire image by a random degree ($±5^\circ$) to simulate non-aligned document scans or hand-held photos.
• —Augmentations (Blur): Randomly applies a Gaussian blur to simulate out-of-focus captures or low-resolution sensor output.
• —Augmentations (Noise): Injects digital grain to mimic the sensor noise typically found in low-light photography.