Paint By Relaxation

A much simplified implementation of paper Paint by relaxation, a bit more like Painterly rendering with curved brush strokes of multiple sizes in practice, coursework for Computer Vision course at Cardiff University in 2022.

Coursework Pro-forma

Example Usage

java Relaxation your_image.ppm example-brushes/brushSquare.pgm example-brushes/brushEllipse.pgm 0.01

Compile & Generate Javadoc

javac Relaxation.java
javadoc -d doc ./*.java

Usage Help

Usage:
> java Relaxation -h | --help
> java Relaxation <input_image> <compact_brush> <elongated_brush> <density> [-f] [-t <threads>] [-r <seed>] [-s <scale>] [-n <std>]
Options:
  -f force the program to proceed with an input image at any size
  -t specifies the number of threads to use for painting, 0 (default): use cpu count
  -r specify a random seed for a consistent output
  -s specify a scaling factor for brush images
  -n specify a standard deviation for optional gaussian noise added to the smaller strokes in the painting
  (if a negative value is set for threads, an unsafe multi-threading strategy is used)

The -n option can sometimes add a tiny bit of desired randomness.

Input Image	Painted Output	With Imperfection (-n 100)

Stylized Animation (created via Std-Raytracer):

Animation.mp4

The Painting Process

Optimistic thread locking is used for safe multi-threading painting, unsafe version does not check the pixel version numbers

1. Load inputs: target image, two brush masks (compact + elongated), and a stroke density parameter.
2. Build an image pyramid of the target (repeated 2× downsampling).
3. Compute Sobel edge magnitude across pyramid levels (per RGB channel), then combine into one full-resolution edge-strength map.
4. Compute Sobel edge direction (orientation map) from a smoothed grayscale version of the target.
5. Precompute brush variants: a small set of discrete sizes and discrete orientations (scaled + rotated masks).
6. Compute a multi-scale DoG (Difference of Gaussians) map and a thresholded map for region classification.
7. Paint coarse-to-fine: for each scale, sample many candidate stroke positions; choose stroke size from edge strength (stronger edges → smaller), orientation from Sobel direction, colour as mean RGB under the mask; apply a greedy accept/reject test (only keep strokes that improve similarity).
8. Second pass: repeat painting using elongated strokes in regions indicated by the DoG threshold, typically with higher density.
9. Deal with unpainted areas: region growing to form patches of unpainted areas, fill them with average colour.

From left-to-right: Five level of stroke sizes, right-most image is the final output

Format Java Code via Clang-Format

find . -name "*.java" -exec clang-format -i {} +