Abstract
The aim of image segmentation in imaging science is to solve the problem of partitioning an image into smaller disjoint homogeneous regions that share similar attributes. This paper investigates an improved Level Set Method (LSM) based on the Chan-Vese (C-V) model, incorporating an initialization mask for vector image segmentation across multiple color spaces. The final segmentation is achieved through a simple labeling scheme. Experimental results demonstrate that the optimized C-V model provides faster, more robust, and more accurate segmentation in RGB, CIE XYZ, and YCbCr color spaces. Conversely, the model exhibits significant errors in HSV and CIE Lab* color models, and produces poor reliability in the h1h2h3 space.
Introduction
Image segmentation is a fundamental task in computer vision, essential for isolating objects of interest from their backgrounds. The Level Set Method (LSM) has emerged as a powerful technique for this purpose due to its ability to handle complex topological changes automatically. However, standard LSMs, such as the popular Chan-Vese model, often suffer from high sensitivity to initial contour placement and inconsistent performance across different color representations. This research addresses these challenges by introducing a structured initialization mask and conducting a comparative study across various color spaces to identify the most effective environments for robust segmentation.
Methodology
The proposed approach builds upon the Mumford-Shah framework, specifically the vector-valued Chan-Vese model. This model treats segmentation as an energy minimization problem, seeking a piecewise constant approximation of the image across multiple channels.
System Architecture
The segmentation pipeline involves four primary stages: inputting the color image, converting it into the target color space, applying the C-V model with the proposed masking grids, and finally generating the segmented image.
Initialization Masking Grids
To mitigate the "initialization sensitivity" of traditional LSMs, the authors propose the use of structured masking grids. Instead of manual or random initialization, diamond-shaped grids are used to generate an initial level set function ($u_0$) via a Euclidean distance transform. This structured start ensures faster convergence and prevents the contour from becoming trapped in local minima during the evolution process.
Color Space Comparative Study
A key contribution of this work is the evaluation of the refined C-V model across six distinct color spaces: RGB, CIE XYZ, YCbCr, YIQ, HSV, and CIE Lab*. While RGB is the standard, its channels are highly correlated, which can hinder the distinction of certain objects. The researchers hypothesized that other spaces might offer better decorrelation for segmentation tasks.
Key Results
- High Performance: The model demonstrated excellent adaptability and noise robustness in RGB, CIE XYZ, and YCbCr spaces. In these environments, the algorithm successfully extracted complex objects, such as a waterlily, with minimal variation in results.
- Performance Failures: Surprisingly, the model struggled significantly in HSV and CIE La*b spaces. The piecewise constant assumption of the Chan-Vese model appears to conflict with the non-linear data distribution in these transformed spaces, leading to fragmented or inaccurate segmentations. The h1h2h3 space also produced poor reliability and accuracy.
- Efficiency: The use of the initialization mask significantly reduced the number of iterations required for convergence compared to standard C-V implementations, leading to faster overall processing times.
Conclusion
The integration of a structured initialization mask into the Chan-Vese level set method provides a robust solution for color image segmentation. While the choice of color space is critical—with RGB, XYZ, and YCbCr being the most reliable—the proposed method offers a significant improvement in both speed and accuracy. Future research will focus on developing more objective evaluation metrics and investigating the specific mathematical causes for the model's failure in non-linear color spaces like HSV and Lab*.
Summary updated