Detection of floating mines in infrared sequences by multiscale geometric filtering

Dominique Florins, Antoine ManzaneraProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2012 · DOI

TL;DR

This paper presents a novel algorithm for detecting floating mines in infrared video sequences by analyzing their geometric properties. By combining multiscale differential geometry with spatiotemporal filtering, the method effectively identifies elliptical shapes against the sea surface, offering a promising tool for maritime security.

What Problem Does This Paper Address?

The automatic detection of floating mines in maritime environments is a critical challenge for naval security. Traditional methods often struggle with the mines' limited visibility, variable appearance due to waves and weather, and the cost-effectiveness of passive sensing. This paper tackles the problem of reliably detecting these mines in infrared imagery, where they typically appear as bright elliptical areas due to emissivity differences with the sea at low incidence angles.

Methods

The core of the proposed algorithm lies in a multiscale differential geometry approach. The authors leverage the Hessian matrix, which captures the second-order derivatives (curvatures) of the image intensity. By analyzing the eigenvalues of this matrix across different scales, they can characterize the shape and contrast of potential mine candidates.

Key steps include:

  • Scale Selection: Determining relevant scales based on the camera's geometry and expected mine sizes.
  • Spatiotemporal Filtering:
    • Temporal Smoothing: An exponential moving average is used to reduce noise and fluctuations caused by waves.
  • Spatial Contrast Enhancement: A morphological contrast filter enhances the visibility of features at selected scales.
  • Geometrical Detection:
    • Ellipticity Measure: A parameter is defined based on the eigenvalues of the Hessian matrix to quantify the aspect ratio (eccentricity) of elliptical regions.
  • Contrast Measure: The Frobenius norm of the Hessian matrix is used to measure the local contrast.
  • Mine-likeliness: A combined measure is derived from the ellipticity and contrast, highlighting bright elliptical regions.
  • Temporal Stability: Detections are retained only if they are consistently present over a defined temporal window, using a dilation-based approach to account for slight displacements.
  • Tracking: A moving average of stable detections is computed, and the boundary of this smoothed region is used to track the mine, even through temporary occlusions.

Datasets

The authors evaluated their algorithm on a dataset of 10 infrared video sequences of floating test targets. These sequences were captured under various environmental conditions, with some containing multiple mines and others exhibiting significant swell that obscured the mines. The dataset includes both mid-infrared (3-5µm) and long-infrared (8-12µm) bands. Ground truth was provided as binary images indicating mine locations.

Key Insights and Ideas

  • Geometry as the Primary Cue: The paper strongly emphasizes that the geometric characteristics of floating mines (specifically, their elliptical shape in infrared imagery) are the most robust features for detection, rather than relying solely on motion.
  • Multiscale Analysis is Crucial: Analyzing image features across multiple scales is essential to capture mines of varying apparent sizes and to robustly estimate their geometric properties.
  • Spatiotemporal Filtering for Robustness: Combining spatial contrast enhancement with temporal smoothing effectively filters out noise from waves and enhances the mine's signature, while temporal stability ensures that detections are not fleeting false alarms.
  • Infrared Advantage: The paper highlights how emissivity differences in infrared imagery at low incidence angles create a strong contrast between mines and the sea, making it a suitable sensor for this task.

Key Results

The algorithm demonstrated strong performance across the 10 sequences. The paper reports:

  • Precision: Ranging from 25.95% to 100.00% across different sequences and validity thresholds.
  • Recall: Ranging from 21.71% to 100.00% across different sequences and validity thresholds.

The authors note that recall can be limited by the visibility of the mine in the sequence, especially when obscured by waves. The "Ship" sequence, captured from a frigate, is highlighted as being particularly challenging and realistic.

Limitations and Future Work

  • Recall Limitations: The recall is inherently limited by the mine's visibility in the infrared sequence. When a mine is significantly obscured by waves, the algorithm cannot detect it, impacting the recall metric.
  • Ground Truth Construction: For evaluation, the ground truth masks were made to always show the mine, even when hidden, to allow for tracking evaluation. This approach influences the recall calculation.
  • Future Work: The authors suggest that this method can serve as a first step in automated detection systems, providing regions of interest for higher-level classification. Further improvements could involve real-time embedded implementation and evaluation of the on-board system. They also mention the potential for using this approach in fully automated detection systems.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{florins2012detection,
  author = {Dominique Florins and Antoine Manzanera},
  title = {{Detection of floating mines in infrared sequences by multiscale geometric filtering}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2012},
  doi = {10.1117/12.918420},
}
Figure