Description

The GR32.Blur.SelectiveGaussian unit implements edge-preserving (bilateral-like) Gaussian blurring for 32-bit ARGB bitmaps.

While standard Gaussian blurring (Blur32) averages pixels across a spatial neighborhood regardless of color boundaries, causing sharp edges and fine details to become uniformly blurry, selective Gaussian blurring filters pixels conditionally based on color similarity.

---

Core Concept: Spatial Radius vs. Color Delta

Selective Gaussian blurring evaluates neighbor pixels within a spatial neighborhood (defined by Radius) but includes a neighbor pixel in the weighted Gaussian average only if its color intensity differs from the center reference pixel by no more than a threshold value (Delta).

1. Spatial Blur Radius (Radius)

The Radius parameter defines the spatial neighborhood (in pixels) evaluated around each pixel. Larger radius values allow the filter to consider more surrounding pixels for smoothing, but increase computational complexity.

2. Color Intensity Threshold (Delta)

The Delta parameter (typically $1\dots 255$) specifies the maximum allowable color channel difference between neighbor pixels and the center reference pixel:

$$|C_{\text{sample}}-C_{\text{ref}}|\le \text{Delta}$$

• Small Delta (e.g. 5..20): Only pixels with nearly identical colors are averaged. Sharp edges, object outlines, and high-contrast details are preserved completely intact while subtle noise, gradients, and flat surface textures are smoothly blurred.
• Large Delta ($\ge 255$): All pixels within the spatial radius satisfy the threshold, causing the selective blur to behave like standard spatial Gaussian blurring.

---

Practical Applications

• JPEG Artifact & Noise Reduction: Removes blocking, ringing, and mosquito noise from compressed photos while maintaining crisp object edges and text boundaries.
• Skin & Surface Smoothing: Smooths skin tones and surface textures in photographic retouching without blurring eyes, hair, or sharp image contours.
• Cartoon / Posterization Preparation: Flattening color variations within solid regions prior to vectorization or edge detection.

Before	After

Removing JPEG artifacts with Selective Blur. Radius: 5.0, Threshold: 15

---

Gamma Awareness & Channel Processing

• Gamma-Aware Blurring: GammaSelectiveGaussianBlur32 performs color thresholding and kernel weighting in linear light space (using GR32_Gamma), eliminating dark halo artifacts along high-contrast boundaries.
• Alpha Channel Policy: Selective Gaussian blurring processes Red, Green, and Blue channels. By design, the Alpha channel of the target bitmap is preserved without being blurred.

---

Note on Experimental Implementations

The GR32.Blur.SelectiveGaussian unit contains various internal and experimental reference implementations (such as SIMD-optimized passes, cache-accelerated algorithms, and historical GIMP adaptation functions). Only the public entry points (SelectiveGaussianBlur32 and GammaSelectiveGaussianBlur32) are intended for public use.

See also

• GR32.Blur
• GR32_Gamma