Kullback-Leibler Divergence-Based Fuzzy $C$-Means Clustering Incorporating Morphological Reconstruction and Wavelet Frames for Image Segmentation

Cong Wang, Witold Pedrycz, Zhiwu Li, MengChu Zhou

TL;DR

This paper introduces a novel image segmentation algorithm, KLDFCM, that balances high accuracy with computational efficiency. By integrating morphological reconstruction for noise reduction and wavelet frames for feature extraction, alongside a Kullback-Leibler divergence term to refine pixel memberships, KLDFCM achieves superior segmentation results while being faster than many existing methods.

What Problem Does This Paper Address?

Image segmentation, the process of partitioning an image into meaningful regions, is a fundamental task in computer vision. The Fuzzy C-Means (FCM) algorithm is a popular choice due to its simplicity, but it struggles with noisy images and can be computationally expensive when incorporating spatial information to improve robustness. Existing methods that enhance FCM's robustness often increase computational costs significantly, creating a trade-off between performance and speed. This paper aims to bridge this gap by developing an FCM variant that is both robust to noise and computationally efficient.

Methods

The proposed KLDFCM algorithm introduces several key innovations:

  • Morphological Reconstruction (MR) for Pre-processing: The input image is first filtered using MR. This step is crucial for enhancing robustness to noise and preparing the image data for fuzzy clustering by preserving image details. MR is presented as a more effective and less parameter-dependent alternative to traditional filters.
  • Tight Wavelet Frame Transform for Feature Extraction: Instead of directly using image pixels, the algorithm employs a tight wavelet frame transform to decompose both the original and MR-filtered images. This creates feature sets that capture richer information, akin to using a kernel function to map data to a new feature space where it's easier to analyze underlying features and noise.
  • Kullback-Leibler (KL) Divergence in the Objective Function: A KL divergence term is incorporated into the FCM objective function. This term encourages the membership degrees of a pixel to be similar to those of its neighbors, leading to a more refined and suitable membership partition. A significant advantage is that this approach eliminates the need for the fuzzification coefficient (m) typically found in FCM, simplifying parameter tuning.
  • Reconstruction of Segmented Feature Set: After obtaining the initial segmentation, a reconstruction process is applied to the segmented feature set. This step further refines the segmentation by minimizing the inverse of the modified objective function. Abnormal features are then corrected by reassigning them to the closest prototype.
  • Wavelet Frame Reconstruction: Finally, the segmented image is reconstructed using the inverse wavelet frame transform.

Datasets

The paper evaluates the proposed KLDFCM algorithm on:

  • Synthetic images: Both with and without ground truth, corrupted by mixed Gaussian and impulse noise.
  • Medical images: Simulated Magnetic Resonance Images (MRIs) from the BrianWeb database, with specified noise and intensity non-uniformity levels.
  • Color images:
    • Images from the Berkeley Segmentation Dataset (BSDS300).
  • Real-world images from the NASA Earth Observation database (sea ice, snow extent, and chlorophyll concentration).

Key Insights and Ideas

  • MR as a Smart Pre-processor: Morphological reconstruction proves to be an effective way to introduce spatial information and enhance noise immunity without the computational burden of iterative neighborhood calculations common in other FCM variants.
  • Wavelet Frames as Implicit Kernels: Using wavelet frames to generate feature sets acts as a form of implicit kernelization, transforming the data into a space where features are more discriminative and noise is better handled.
  • KL Divergence for Smoother Boundaries: The KL divergence term is a clever way to enforce local consistency in pixel memberships, leading to more coherent and accurate segmentation boundaries, especially in noisy regions.
  • Simplified Parameter Tuning: Eliminating the fuzzification coefficient (m) is a practical benefit, making the algorithm easier to use.
  • Reconstruction for Refinement: The two-stage reconstruction process (feature set and then image) is key to further improving segmentation accuracy by correcting artifacts introduced during the initial segmentation.

Key Results

  • Superior Segmentation Accuracy: KLDFCM consistently achieved higher segmentation accuracy (SA) across synthetic, medical, and color images compared to eight other state-of-the-art FCM-related algorithms. For instance, on synthetic images, KLDFCM achieved SA up to 99.9630%.
  • Improved Robustness: The algorithm demonstrated strong performance in the presence of high-intensity mixed Gaussian and impulse noise, effectively removing noise while preserving image details and clear edges.
  • Computational Efficiency: While not the absolute fastest, KLDFCM offered a good balance between performance and speed. It was significantly faster than several comparative algorithms and, despite being slightly slower than FRFCM (which uses gray level histograms), its superior segmentation performance made it more practical.
  • Effective for Color Images: The method was successfully extended to color image segmentation by applying MR and wavelet transforms channel-wise, outperforming other algorithms in this domain as well.

Limitations and Future Work

  • Parameter Sensitivity: While the fuzzification coefficient is removed, the parameters $\alpha$ and $\beta$ for the KL divergence term still require careful selection. The paper provides experimental guidance for their selection but notes that automatic parameter tuning could be a future direction.
  • Automatic Cluster Number Selection: The number of clusters ($c$) is assumed to be known, which is a common limitation in many clustering algorithms. Future work could explore automatic methods for determining the optimal number of clusters.
  • Expanded Application Domains: The authors suggest that the algorithm's applicability could be further expanded to non-flat domains like remote sensing, computer networks, ecological systems, and transportation networks.

Summary updated

Abstract

Although spatial information of images usually enhance the robustness of the Fuzzy C-Means (FCM) algorithm, it greatly increases the computational costs for image segmentation. To achieve a sound trade-off between the segmentation performance and the speed of clustering, we come up with a Kullback-Leibler (KL) divergence-based FCM algorithm by incorporating a tight wavelet frame transform and a morphological reconstruction operation. To enhance FCM's robustness, an observed image is first filtered by using the morphological reconstruction. A tight wavelet frame system is employed to decompose the observed and filtered images so as to form their feature sets. Considering these feature sets as data of clustering, an modified FCM algorithm is proposed, which introduces a KL divergence term in the partition matrix into its objective function. The KL divergence term aims to make membership degrees of each image pixel closer to those of its neighbors, which brings that the membership partition becomes more suitable and the parameter setting of FCM becomes simplified. On the basis of the obtained partition matrix and prototypes, the segmented feature set is reconstructed by minimizing the inverse process of the modified objective function. To modify abnormal features produced in the reconstruction process, each reconstructed feature is reassigned to the closest prototype. As a result, the segmentation accuracy of KL divergence-based FCM is further improved. What's more, the segmented image is reconstructed by using a tight wavelet frame reconstruction operation. Finally, supporting experiments coping with synthetic, medical and color images are reported. Experimental results exhibit that the proposed algorithm works well and comes with better segmentation performance than other comparative algorithms. Moreover, the proposed algorithm requires less time than most of the FCM-related algorithms.
BibTeX
@article{wang2020kullback,
  author = {Cong Wang and Witold Pedrycz and Zhiwu Li and MengChu Zhou},
  title = {{Kullback-Leibler Divergence-Based Fuzzy $C$-Means Clustering Incorporating Morphological Reconstruction and Wavelet Frames for Image Segmentation}},
  journal = {IEEE Trans. Cybernetics},
  year = {2020},
}
Figure