Pendekatan Denoising Hibrid untuk Pengenalan Aksara Jawa: Upaya Integrasi Teknologi dan Pelestarian Budaya Digital

Abstrak

The digitization of Javanese manuscripts is hindered by image degradation such as bleed-through, stains, blur, and uneven illumination that obscures fine graphemic details and reduces the reliability of optical character recognition (OCR). This challenge is significant for computational heritage preservation, as degraded visual inputs threaten the accuracy and longevity of digital cultural archives. This study applies and evaluates a Structural-Preserving Denoising (SPD) approach designed to retain graphemic contours during noise reduction. Using a comparative-experimental design, six denoising methods (Wavelet, Bilateral, Non-Local Means/NLM, and three hybrids) were tested as independent variables, with PSNR, SSIM, and OCR accuracy as dependent variables. A dataset of 4,632 degraded Javanese character images (64×64 pixels) was processed using Histogram of Oriented Gradients and classified via a cosine-kernel Support Vector Machine. Analyses included comparative evaluation, error analysis, and Pearson correlation. Results show that Bilateral Filtering achieved the best balance of structural preservation and recognition performance (SSIM 0.9108; PSNR 29.85 dB, accuracy 96.97%). The strong correlation between visual quality and OCR accuracy indicates that maintaining graphemic structure is crucial for robust recognition. The study contributes a theoretical SPD framework and a practical preprocessing pipeline applicable to the digital preservation of traditional scripts.

Identifier

• OpenAlex: https://openalex.org/W7124768884
• DOI: https://doi.org/10.29408/edumatic.v9i3.32376