Novel model using kernel function and local intensity information for noise image segmentation

Gang Li, Haifang Li, Ling ZhangTsinghua 2018 · DOI

Abstract

It remains a challenging task to segment images that are distorted by noise and intensity inhomogeneity. To overcome these problems, this paper presents a novel region-based active contour model based on local intensity information and a kernel metric. By introducing intensity information about the local region, the proposed model can accurately segment images with intensity inhomogeneity. To enhance the model's robustness to noise and outliers, we introduce a kernel metric as its objective functional. To more accurately detect boundaries, we apply convex optimization to this new model, which uses a weighted total-variation norm given by an edge indicator function. Lastly, we use the split Bregman iteration method to obtain the numerical solution. Experiments on synthetic and real images demonstrate significant improvements in efficiency and accuracy compared with currently popular methods.

Introduction: The Challenge of Noise and Inhomogeneity

Image segmentation—the process of partitioning an image into meaningful regions—is a cornerstone of computer vision. However, two "real-world" factors often break standard algorithms:

  1. Noise: Random variations in pixel values (like salt-and-pepper or Gaussian noise) can trick models into seeing boundaries where none exist.
  2. Intensity Inhomogeneity: Uneven lighting or shading across an image means that the "target" object might have different brightness levels in different areas, making global thresholding impossible.

Classic models like the Chan-Vese (CV) model assume that an image is "piecewise constant" (uniform brightness in each region) and use standard Euclidean distance to measure pixel similarity. This makes them fragile in the face of noise and shading. Furthermore, many existing solutions are computationally expensive or sensitive to initial contour placement.

Methodology: Kernel Metrics and Local Information

The proposed Local Piecewise Model based on a Kernel Function introduces several key innovations to address these limitations:

  • The Kernel Metric: Instead of using Euclidean distance, the model uses a Gaussian radial-basis function (a "kernel"). This de-emphasizes outliers; if a pixel's intensity is vastly different from its neighbors (like a noise spike), the kernel function gives it a very low weight, effectively ignoring it.
  • Local Intensity Information: The model calculates "fitting centers" based on local neighborhoods rather than the whole image. This allows the boundary to adapt to local shading and intensity changes.
  • Global Convex Optimization: To avoid getting stuck in "local minima" (suboptimal solutions), the authors reformulate the energy functional into a convex form. This ensures that the algorithm finds the global optimum regardless of where the initial contour is placed.
  • Split Bregman Method: This high-efficiency numerical scheme allows the model to converge in a fraction of the time required by standard gradient descent methods.
3D surface plot (i) showing the intensity profile of a noisy 'C' shape, illustrating the complex intensity landscape the model must navigate.

Experimental Evaluation and Results

The model was evaluated using a mix of synthetic images with blurred boundaries, medical cell images, and the Berkeley Segmentation Dataset. To test robustness, data was artificially injected with Salt-and-pepper, Gaussian, and Speckle noise.

Segmentation results across various datasets: (a-b) cell images, (c-d) synthetic shapes with noise, and (e-f) a heavily noise-corrupted 'C' shape.

Key Performance Metrics

  • Speed: In tests on infrared images, the proposed method converged in just 1 to 3 iterations (approx. 0.28 seconds), while the competing RSF model required 120 iterations (approx. 10.7 seconds).
  • Accuracy: Using the Dice Similarity Coefficient (DSC), the model maintained scores above 0.89 even under heavy speckle noise, significantly outperforming the LIC model which dropped to 0.59 under similar conditions.
  • Initialization Independence: Because the model is convex, the "active contour" can be initialized anywhere—inside, outside, or across the object—and still converge to the correct boundary.

Conclusion and Future Work

By moving to a kernel-based feature space, the model becomes naturally "blind" to noise while remaining sharp on actual object boundaries. While the paper demonstrates excellent results, future work will explore extending the framework to multi-phase segmentation (identifying multiple object types) and 3D application for medical imaging like MRI or CT scans.

Summary updated

Abstract

It remains a challenging task to segment images that are distorted by noise and intensity inhomogeneity. To overcome these problems, in this paper, we present a novel region-based active contour model based on local intensity information and a kernel metric. By introducing intensity information about the local region, the proposed model can accurately segment images with intensity inhomogeneity. To enhance the model's robustness to noise and outliers, we introduce a kernel metric as its objective functional. To more accurately detect boundaries, we apply convex optimization to this new model, which uses a weighted total-variation norm given by an edge indicator function. Lastly, we use the split Bregman iteration method to obtain the numerical solution. We conducted an extensive series of experiments on both synthetic and real images to evaluate our proposed method, and the results demonstrate significant improvements in terms of efficiency and accuracy, compared with the performance of currently popular methods.
BibTeX
@inproceedings{li2018novel,
  author = {Gang Li and Haifang Li and Ling Zhang},
  title = {{Novel model using kernel function and local intensity information for noise image segmentation}},
  booktitle = {Tsinghua},
  year = {2018},
  doi = {10.26599/tst.2018.9010001},
}
Figure