Randomized algorithms for statistical image analysis and site percolation on square lattices

We propose a novel probabilistic method for detection of objects in noisy images. The method uses results from percolation and random graph theories. We present an algorithm that allows to detect objects of unknown shapes in the presence of random noise. The algorithm has linear complexity and exponential accuracy and is appropriate for real-time systems. We prove results on consistency and algorithmic complexity of our procedure.

💡 Research Summary

The paper introduces a novel statistical framework for detecting objects in noisy digital images by leveraging concepts from site percolation theory and random graph algorithms. The authors model a grayscale image as an N × N array of observations Y₍ᵢⱼ₎ = I₍ᵢⱼ₎ + σ ε₍ᵢⱼ₎, where the true underlying image I₍ᵢⱼ₎ is binary (1 for object pixels, 0 for background) and the noise ε₍ᵢⱼ₎ are independent, identically distributed according to a known distribution F. No assumptions of smoothness, symmetry, or even finite moments are required for the noise; only its distribution function is needed.

The central idea is to apply a global threshold θ(N) to the noisy grayscale values, converting each pixel to a binary value: pixels with Y₍ᵢⱼ₎ ≥ θ(N) become black (1), the rest become white (0). The threshold is chosen so that two inequalities hold simultaneously:

1. The probability that a background pixel is mistakenly turned black, 1 − F(θ/σ), is smaller than the critical site‑percolation probability p_c on the square lattice.
2. The probability that a true object pixel remains black, 1 − F((θ − 1)/σ), is larger than p_c.

When these conditions are satisfied, the transformed binary image induces a random subgraph G_N of the Z² lattice where black vertices correspond to “occupied” sites. Because the black‑pixel conversion probability exceeds p_c inside the true object region, that region experiences a super‑critical percolation regime: large connected black clusters are highly likely to appear. Conversely, the background experiences a sub‑critical regime, producing only small, isolated black clusters. This stark contrast in percolation phases provides a statistically powerful discriminator between object and background.

The detection algorithm proceeds as follows:

1. Compute θ(N) (or a suitable approximation) using either a minimization of the squared distance to p_c (Equation 10) or a maximization of the signed difference (Equation 11).
2. Threshold the noisy image at θ(N) to obtain a binary image.
3. Construct a planar graph G_N where vertices are pixels and edges connect orthogonal neighboring pixels of the same color.
4. Identify the connected components (clusters) of black vertices using a linear‑time graph traversal (e.g., BFS or Union‑Find).
5. Evaluate a test statistic such as the size of the largest black cluster or the total number of black vertices. If the statistic exceeds a pre‑specified level, reject the null hypothesis H₀ (no object) in favor of H₁ (object present).

Complexity analysis shows that each step requires O(N²) operations, i.e., linear time in the number of pixels, making the method suitable for real‑time applications. Theoretical results from percolation theory guarantee that, under H₁, the probability of observing a super‑critical black cluster grows as 1 − exp(−c N²) for some constant c > 0, yielding exponential accuracy as the image size increases. Under H₀, the probability of a false detection can be bounded by a user‑chosen significance level α(N), which can be tuned by adjusting θ(N) and the stopping rule.

A notable strength of the approach is its robustness to the noise distribution. The only requirement is that the cumulative distribution function F be known, allowing the method to accommodate heavy‑tailed, asymmetric, or even discrete noise models. The authors also discuss how the threshold can be adaptively chosen to maximize the separation between the two percolation probabilities, ensuring that the transformed image is as far from the critical point as possible.

The paper includes a simulation study and a real‑data example (e.g., medical imaging and remote sensing) demonstrating that the proposed algorithm outperforms traditional wavelet‑based or filter‑based detectors, especially when objects have irregular, non‑convex, or fragmented shapes. The method does not rely on any smoothness or shape priors, which is a significant advantage over many existing techniques that require preprocessing or reconstruction before detection.

In summary, the authors present a theoretically grounded, computationally efficient, and practically versatile algorithm for object detection in noisy images. By translating the detection problem into a percolation‑phase discrimination task, they achieve linear‑time performance with exponential convergence guarantees, opening new avenues for fast, reliable image analysis in fields such as medical diagnostics, satellite imagery, and industrial inspection. Future work suggested includes extensions to multi‑object detection, incorporation of color information, and hardware implementations for ultra‑fast processing.