Abstract This work presents a novel approach to photo stylized rendering. By leveraging the Simple Linear Iterative Clustering (SLIC) method, the stylization process is significantly accelerated while maintaining high-quality artistic effects. Regions are constructed using SLIC, a technique traditionally used in image segmentation. We extract the mean color value for each region and render the final image. Experimental results demonstrate satisfying visual outcomes and high computational efficiency.
Introduction
The field of Non-Photorealistic Rendering (NPR) aims to transform standard photographs into artistic styles, such as oil paintings or watercolors. Watercolor, or "aquarelle," is particularly difficult to simulate due to the complex physical interactions between pigments and paper, which create a distinct "piece-wise" appearance. Existing methods often struggle with high computational costs or fail to preserve the structural integrity of the original image. This paper addresses these gaps by repurposing the SLIC superpixel algorithm to generate regular, compact regions that mimic intentional brushstrokes.
Methodology
The core of our approach is the adaptation of the SLIC (Simple Linear Iterative Clustering) algorithm for artistic rendering. SLIC clusters pixels in a five-dimensional space: three channels for color (CIELAB space) and two for spatial coordinates (x, y). This ensures that the resulting "superpixels" are both color-consistent and spatially compact.
The process involves four key steps:
- Seed Initialization: The image is divided into a grid of $M$ initial cluster centers.
- Gradient-Based Refinement: Seeds are moved to the lowest gradient position in a $3 \times 3$ neighborhood to avoid placing centers on sharp edges or noise.
- Iterative Optimization: Pixels are assigned to the nearest cluster center using a weighted distance formula balancing color and spatial proximity.
- Aquarelle Rendering: Once regions are defined, every pixel within a superpixel is replaced with the mean RGB intensity of that region. This creates the stylized, segmented look characteristic of watercolor art.
Key Results and Discussion
The method was evaluated using the MSRA dataset, focusing on natural subjects like plants and flowers. The algorithm successfully simplified complex details into artistic "aquarelle" blocks while maintaining the original's color gradations and structural saliency.
Performance and Efficiency
A major advantage of this method is its speed. Because SLIC is a linear clustering method, it is significantly faster than traditional segmentation-based NPR techniques. Mean processing times per image were recorded as follows:
- 125 superpixels: ~0.21 seconds
- 250 superpixels: ~0.30 seconds
- 500 superpixels: ~0.50 seconds
Experimental results suggest a "sweet spot" of 250 to 375 superpixels, which provides the best balance between artistic abstraction and detail preservation.
Conclusion and Future Work
By treating superpixels as artistic primitives, we have developed a fast, automated pipeline for watercolor stylization. While the method performs exceptionally well on high-contrast natural scenery, future work will focus on improving results for low-contrast images and complex human subjects. Additionally, incorporating physical simulations of "wet-on-wet" bleeding or paper texture could further enhance the realism of the digital aquarelle effect.
Summary updated