A Two-Stage Evolutionary Fuzzy Clustering Framework for Noisy Image Segmentation

Licheng Jiao, Mengxuan Zhang, Fang Liu, Wenping Ma, Lingling Li

Introduction

Image segmentation—the process of partitioning an image into meaningful regions—is a cornerstone of computer vision. A primary challenge in this field is the sensitivity of traditional algorithms, such as Fuzzy C-Means (FCM), to noise. Because FCM often treats pixels in isolation, it struggles with noisy data. While researchers frequently introduce "spatial information" to smooth out noise, this creates a difficult trade-off: excessive smoothing destroys fine details and sharp edges, while insufficient smoothing leaves noise intact. Finding the "sweet spot" typically requires manual parameter tuning for every image.

The Two-Stage Evolutionary Fuzzy Clustering (TEFC) framework automates this balance. By identifying an optimal trade-off on difficult "inhomogeneous" regions and applying that logic to the entire image via an adaptive evolutionary algorithm, it significantly outperforms traditional methods on both synthetic and complex real-world imagery.

The TEFC Framework

The authors propose a two-stage approach designed to find the optimal balance between noise suppression and detail preservation automatically.

The TEFC framework overview, illustrating the transition from multi-objective optimization to final segmentation.

Stage 1: Multi-Objective Optimization

Instead of processing the entire image—which is computationally expensive—the algorithm samples "inhomogeneous regions" like edges or textures where noise and detail are hardest to distinguish. It treats segmentation as a multi-objective problem with two conflicting goals: preserving image details and restraining noise. Using the MOEA/D evolutionary algorithm, it generates a "Pareto front" of possible solutions. The framework automatically selects the solution in the "knee region" of this front—the point representing the best compromise between the two objectives.

Stage 2: Fuzzy Clustering Segmentation

The optimal trade-off found in Stage 1 is then applied to the entire image. To avoid "local optima" (mathematical traps where an algorithm settles for a sub-optimal answer), the framework employs an adaptive evolutionary fuzzy clustering algorithm. This stage evolves cluster centers over several generations to reach a globally optimal segmentation.

Intelligent Sampling for Efficiency

A key innovation of TEFC is its smart sampling strategy. By focusing heavy-duty multi-objective optimization only on inhomogeneous regions, the algorithm saves significant computation time while focusing its "intelligence" exactly where it is needed most.

Detailed view of the inhomogeneity measure based sampling strategy.

This sampling strategy ensures that the parameters tuned in Stage 1 are specifically optimized for the most challenging parts of the image, ensuring that edges are preserved even in high-noise environments.

Experimental Results and Insights

The TEFC framework was evaluated against several state-of-the-art methods, including various FCM variants and Ncut, across diverse datasets:

  • Synthetic Images: TEFC maintained high performance (measured by Accurate Rate and Adjusted Rand Index) even as Gaussian and Salt & Pepper noise levels increased.
  • Natural Images: Benchmarks from the Berkeley Segmentation Dataset (BSD) and VOC2012 confirmed the framework's versatility.
  • SAR (Synthetic Aperture Radar) Images: On the Lake SAR dataset, which is notoriously difficult due to "speckle" noise, TEFC achieved a Partition Coefficient (Vpc) of 0.9547, significantly higher than its competitors.

Qualitatively, TEFC produces results with cleaner edges and smoother regions. Unlike traditional methods that often leave noise "spots" or blur thin structures, TEFC successfully navigates the noise-detail trade-off.

Conclusion and Future Work

TEFC demonstrates that evolutionary algorithms can effectively automate the complex parameter tuning required for robust image segmentation. While the current framework primarily utilizes gray-level intensity and spatial position, future work may incorporate color and texture information to further enhance performance in complex natural scenes. Additionally, making the sampling parameters self-adaptive based on image complexity could further improve the framework's efficiency.

Summary updated

Abstract

This article presents a two-stage evolutionary fuzzy clustering framework for noisy image segmentation. It is a bi-stage system comprising a multi-objective optimization stage and a fuzzy clustering segmentation stage. In the multi-objective optimization stage, the fuzzy clustering on pixels in inhomogeneous regions is converted into a multi-objective problem, which can preserve image details while restraining noise. The multi-objective problem is decomposed into several sub-problems by the Tchebycheff approach. A trade-off can be obtained by optimizing these sub-problems simultaneously. In the fuzzy clustering segmentation stage, fuzzy clustering with the trade-off between preserving image details and restraining noise is performed on the whole observed image. To deal with this fuzzy clustering problem, an adaptive evolutionary fuzzy clustering algorithm with spatial information is proposed. Experiment results on synthetic and real images illustrate the effectiveness of the proposed framework for noisy image segmentation.
BibTeX
@article{jiao2020two,
  author = {Licheng Jiao and Mengxuan Zhang and Fang Liu and Wenping Ma and Lingling Li},
  title = {{A Two-Stage Evolutionary Fuzzy Clustering Framework for Noisy Image Segmentation}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3029773},
}
Figure