Introduction
Detecting small targets in infrared (IR) imagery is a critical component of modern search-and-track systems. However, this task is notoriously difficult due to the nature of the targets themselves. In extreme environments—characterized by low illumination or high background clutter—targets often consist of only a few pixels, lacking distinct shape, texture, or color. This paper introduces a robust detection method that mimics the Human Visual System (HVS) to effectively find these "needles in haystacks."
Traditional algorithms frequently struggle with a low Signal-to-Clutter Ratio (SCR), where the target is easily "submerged" in background noise or complex structures like clouds and terrain. The challenge is twofold: suppressing the background (which may have higher intensity than the target) and enhancing the target signal without triggering a flood of false alarms.
Methodology
The proposed framework utilizes a dual-feature approach, combining spatial gray-level analysis with frequency-domain saliency extraction to create a final saliency map.
1. Spatial Gray Level-Based Feature Map (S1)
To isolate targets from complex background textures, the researchers developed a Modified White Top-Hat (MRWTH) filter. Unlike standard top-hat filters that often fail against irregular background edges, this method uses multi-directional linear structuring elements (eight directions at $\pi/8$ intervals). This is followed by a Multi-scale Binomial Kernel, which applies a weighted map derived from the Laplacian operator to enlarge the contrast between a target pixel and its immediate neighbors.
2. Saliency Extraction-Based Feature Map (S2)
This stage focuses on the "pop-out" characteristics of small targets. It employs Bilateral Filtering, which preserves edges while smoothing noise, followed by a Difference of Bilateral (DoB) calculation. By finding the difference between a bilateral-filtered image and a neighborhood average, the system extracts high-frequency saliency features, effectively acting as a band-pass filter.
Feature Fusion
The final saliency map is the product of these two maps ($S = S1 \cdot S2$). This multiplication ensures that a region is only flagged as a target if it exhibits both high local contrast (spatial) and high saliency (frequency), acting as a logical "AND" gate to drastically reduce false positives.
Experimental Evaluation
The algorithm was evaluated using both real-world IR images captured by detectors and simulated images created using a "random midpoint displacement" method to generate realistic, irregular cloud backgrounds.
These datasets allowed the researchers to test the algorithm against weak targets distributed randomly in highly complex scenes, mimicking the directional sensitivity and contrast-based attention of the human eye.
Results and Performance
The proposed method was compared against four established algorithms: Modified Top-hat (MT), Fast Saliency (FS), Multi-Patch Contrast Method (MPCM), and Relative Local Contrast Method (RLCM).
Key performance metrics included:
- SCR Gain (SCRG): The proposed method consistently achieved the highest SCRG. In one real-world scenario, it achieved an SCRG of 26.89, nearly triple that of the next best method (MPCM at 9.27).
- Background Suppression Factor (BSF): The method excelled at cleaning up the image, reaching a BSF of 63.74 in the same scenario, compared to MPCM’s 9.51.
- ROC Curves: Receiver Operating Characteristic (ROC) curves demonstrated that the method maintains a high probability of detection even at extremely low false-alarm rates ($10^{-3}$).
Conclusion
By combining spatial analysis for edge preservation and frequency analysis for noise suppression, this dual-feature approach offers a significant advancement in IR target detection. While performance can drop if input images are severely degraded by hardware malfunctions, the method proves highly resilient to natural background interference. Future work will focus on improving robustness in even more inhomogeneous environments and refining the fusion process for extreme clutter.
Summary updated