Higurashifier

Simple image processing script to apply a higurashi background style filter to an image.

Based on rwbyrocks higurashify.

Process

We cut apply a preprocessing step for resizing and/or cropping the image, and then we pass it through a pipeline with 5 steps (every step can be optional).

In order:

Contrast stretching

When provided a white contrast level percentage $w$, we can derive:

$$b = 1 - w$$ $$w_c = w \times 255$$ $$b_c = b \times 255$$ $$s = \frac{255}{w_c - b_c}$$

Then for each RGB channel $c$, we apply the following transformation:

$$c' = clamp((c - b_c) \times s, 0, 255)$$

(Un)Sharpening

We apply an unsharpening mask using a 1D box blur with radius $r = 2$.

When provided a sharpening amount percentage $a$, we apply for each RGB channel $c$ the following transformation:

$$c' = clamp(c + a \times (c - blur(c)), 0, 255)$$

Motion blur

We copy the main image and apply a directional box blur. The user provides as arguments the blur direction angle $a$ (in degrees), the blur radius $r$, and the blur opacity $o$. We can derive:

$$\theta = a \times \frac{\pi}{180}$$ $$dx = cos \theta$$ $$dy = sin \theta$$

For each pixel $(x, y)$, we sample and average around the line

$$(x + k \times dx, y + k \times dx)$$ $$ k \in [-10, 10]$$

Then we alpha blend the blurred result over the image; for each RGB channel $c$ we apply: $$c' = clamp(c \times (1 - o) + blur(c) \times o, 0, 255)$$

Posterization

We apply a simple color quantization over every pixel. The user provides the quantization level $l$ and we derive

$$s = \frac{255}{l - 1}$$

Then for each RGB channel $c$, we apply the following transformation:

$$c' = round(\frac{c}{s}) \times s$$

Edge detection

First, we generate a grayscale image from the original unprocessed pixels. For each RGB value in each pixel in the image:

$$P = 0.30 \times R + 0.59 \times G + 0.11 \times B$$

We apply an edge detection kernel over the grayscale image, any of the following:

• Sobel
• Laplacian
• Roberts
• Prewitt

The user provides a max threshold value $t$ for the edge mask $E$ and invert the output (if we prefer white edges).

$$E' = 255 \hspace{4px} if E > t \hspace{4px} else \hspace{4px} 0$$ $$E_{mask} = 255 - E'$$

Finally we multiply blend the edge mask with the image, the user defines the edge opacity percentage $o$. For each RGB channel $c$:

$$B = \frac{c}{255}$$ $$L = \frac{E_{mask}}{255}$$ $$c' = clamp(255 \times (B \times (1-o) + B \times L \times o), 0, 255)$$

Original image

Result

Parameters:

• aspect: original
• max-height: 1080
• contrast-level: 85.0
• blur-angle: -25.0
• blur-opacity: 0.8
• blur-radius: 9
• edge-method sobel
• edge-threshold: 80.0
• edge-opacity: 0.5
• unsharpening: 1.0
• quantization: 9