Performs image style transfer using classical image processing — no deep learning required.
It works by matching and replacing small image patches between a content image and a style image, transferring texture and tone statistics from one to the other.
Traditional neural style transfer (like Gatys et al., 2015) uses CNN feature maps to blend content and style.
This method achieves a similar effect purely through patch-based texture matching.
- An image’s content = its spatial structure (shapes, edges, object layout)
- An image’s style = its local texture statistics (tone, contrast, roughness)
By replacing each small patch in the content image with a visually similar patch from the style image, we transfer the style while roughly preserving the content layout.
- Local mean and standard deviation capture basic texture statistics, which are essential for visual style (Julesz, 1962).
- By matching these statistics, the algorithm recreates the visual feel of the style image.
- The method is inspired by non-parametric texture synthesis (Efros & Leung, 1999) and statistical texture models (Portilla & Simoncelli, 2000).
Put the content and style image inside the respective directory inside the root images directory. Content image
Resulting image will be saved inside images/output directory.
You can run each operation on a single CPU core or all available cores. The 2nd parameter of the command below define the core count. If it is p, then the function will allocate as many cores are avaiable. If the 2nd parameter = m, it will be a signle core operation.
python main.py p 3 2
The 3rd param = 3 means padding. The 4th param = 2 means stride.