A local Gaussian filter and adaptive morphology as tools for completing partially discontinuous curves

Przemyslaw Spurek, Alena Chaikouskaya, Jacek Tabor, Elzbieta Zajac

Introduction

In fields such as nanotechnology and remote sensing, the identification of continuous structures—like percolation paths in nanomaterials or river networks in satellite imagery—is critical. However, raw imaging data is frequently imperfect. In electron microscopy (EM), metallic nanograins may appear as a series of disjointed dots rather than a solid line due to being partially submerged or scattered. Standard image processing techniques, such as global Gaussian blurring or uniform dilation, often fail to bridge these gaps without introducing significant distortion or merging the signal with background noise. This paper presents a specialized pipeline that utilizes local statistical properties to adaptively reconstruct fragmented curves.

Raw electron microscopy images showing fragmented nanograin paths that require reconstruction.

Methodology

The proposed method consists of a multi-stage pipeline designed to adapt to the local geometry of the image features.

1. Local Gaussian Filtering

Unlike traditional filters that apply a uniform circular kernel, this approach calculates a weighted covariance matrix for the neighborhood of each pixel. This allows the filter to "stretch" and "rotate" its blur to align with the orientation of local curve segments. By subtracting the neighborhood mean before filtering, the method enhances subtle "wire-type" structures while suppressing background noise.

2. Binarization via Cross Entropy Clustering (CEC)

To convert grayscale data into a binary format, the authors employ Cross Entropy Clustering. CEC is demonstrated to be more robust than standard methods like Otsu’s thresholding, particularly when dealing with the noisy and irregular backgrounds typical of nanomaterial imaging.

3. Adaptive Morphology

The core innovation lies in the use of elliptical structural elements for morphological dilation. The dimensions and orientation of these ellipses are derived from the eigenvalues and eigenvectors of the local covariance matrix. If a segment is elongated (indicating a line), the algorithm dilates it aggressively along its primary axis to bridge gaps to adjacent segments. Conversely, circular structures (likely noise) are left unchanged.

The result of adaptive morphology, where elliptical dilation successfully bridges gaps to form a continuous path.

Results and Discussion

The method was evaluated on carbonaceous-nanopalladium (nPd-C) films and satellite imagery of rivers. In EM images, the adaptive pipeline successfully connected large gaps that standard filters could not bridge without creating artifacts. In satellite data, the algorithm correctly identified river paths despite visual obstructions like bridges, preserving the topological integrity of the waterway.

The researchers have implemented these tools as a plugin for ImageJ, making the method accessible for material science and biological research. While the system requires manual tuning of parameters such as neighborhood size and sensitivity, it offers a significant improvement over non-adaptive techniques for completing partially discontinuous curves.

Conclusion

By leveraging local covariance to guide filtering and dilation, this approach provides a robust framework for reconstructing fragmented linear features. Future work may focus on automating parameter selection to further streamline the analysis of complex structural networks.

Summary updated

BibTeX
@inproceedings{spurek2014local,
  author = {Przemyslaw Spurek and Alena Chaikouskaya and Jacek Tabor and Elzbieta Zajac},
  title = {{A local Gaussian filter and adaptive morphology as tools for completing partially discontinuous curves}},
  booktitle = {CISIM},
  year = {2014},
  doi = {10.1007/978-3-662-45237-0\_51},
}
Figure