LOCALLY ADAPTIVE IMAGE ENHANCEMENT WITH RUSCHEWEYH COEFFICIENT-BASED MASKS


Budak S., Çakmak S., Yalçın Tokgöz S.

ASERC 5th INTERNATIONAL CONGRESS ON HEALT, ENGINEERING, ARCHITECTURE and MATHEMATICS, Malatya, Türkiye, 23 - 25 Ekim 2026, ss.128-144, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Malatya
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.128-144
  • İstanbul Gelişim Üniversitesi Adresli: Evet

Özet

Masks built from Ruscheweyh coefficients have been used for image enhancement in earlier work. In these masks a single parameter δ acts on the whole image. When the image is noisy, the same mask that sharpens the edges also amplifies the noise in flat regions. In this paper we propose a locally adaptive mask in which δ is chosen separately at each pixel. The value δ(x,y) lies between δmin and δmax and is controlled by the weight w=v/(v+κ), where v is the local variance. The Koebe function f1, the extremal convex function f2 and the extremal function f3 of bounded turning are used. We show that the adaptive mask can be computed with two fixed convolutions and a pixel-wise linear combination. We also show that it behaves approximately like a sharpener whose gain varies from pixel to pixel. The parameters were selected on 20 validation images of BSDS500. Six standard images and 50 BSDS500 test images were used for testing. The images were degraded by a Gaussian blur with σ=1.5 and by Gaussian noise at three levels. Without noise, the fixed and adaptive masks gave close results. At σn=0.01 the best fixed f2 mask gave no gain over the degraded image (SSIM 0.7190 against 0.7188). The adaptive f2 mask reached an SSIM of 0.7399 and a PSNR of 26.50 dB. On noisy images the adaptive f2 mask performed better than the fixed mask, unsharp masking (USM) and two classical adaptive sharpening methods (Wilcoxon test, p<0.05 in every case). When the same weight map was applied to a Gaussian high-pass component, the SSIM was slightly higher at all three noise levels. Hence the gain comes mainly from the weighting by local structure. The Ruscheweyh framework gives this weighting an analytic form that is controlled by one continuous parameter.

Keywords: Geometric function theory, image enhancement, Ruscheweyh operator, adaptive filter, univalent functions, image processing.