⚡ electrical engineering

Curve Target Extraction from Synthetic Aperture Sonar Image Based on Joint Direction Weighting

This paper proposes a robust method for accurately extracting low-contrast and fractured linear targets, such as underwater cables and pipelines, from synthetic aperture sonar images by utilizing a model that jointly weights pixel gray values with the alignment between local path directions and approximate linear orientations.

Original authors: Zhizhong JIANG, Genxiang LI, Chaoyun YAO, Yaping ZOU

Published 2026-09-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhizhong JIANG, Genxiang LI, Chaoyun YAO, Yaping ZOU

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Beneath the ocean's surface, the world is often a place of profound silence and darkness, where light from the sun cannot reach. To navigate this hidden realm, scientists and engineers rely on sound. Instead of using cameras that need light, they use sonar, a system that sends out sound pulses and listens for the echoes that bounce back. When these pulses are processed with sophisticated techniques known as synthetic aperture sonar, they can create surprisingly detailed images of the seafloor, revealing the shapes of sunken ships, natural formations, and human-made structures like pipelines and cables. These linear objects are vital to our underwater infrastructure, but seeing them clearly is difficult. The images produced are often grainy, dim, and broken by noise, much like trying to read a faint, smudged map in a storm. The sound waves can bounce off the seabed in confusing ways, creating shadows and false reflections that hide the true shape of the target. For years, extracting a clear line from these messy images has been a stubborn problem, as traditional methods often get lost in the static or fail when the line is faint or broken.

A team of researchers from the Hunan University of Information Technology has developed a new way to solve this puzzle, one that pays close attention to the direction a line is moving. Their work focuses on a specific challenge: how to find a continuous path for a cable or pipe when the image is so poor that the line appears to vanish and reappear, or when it is so dim it blends into the background. The researchers realized that while the brightness of a pixel might be unreliable due to noise, the direction of the line itself tends to be consistent. In the real world, a pipeline or a cable does not suddenly twist at a sharp angle from one moment to the next; it flows with a gentle, predictable curve. The team built a method that uses this physical reality as a guide. They designed a system that does not just look for bright spots, but also checks if the path it is tracing is moving in a direction that makes sense based on where it was a moment before.

The process begins by cleaning up the raw image. The researchers first remove the grainy noise and even out the brightness, making the dark areas slightly lighter and the bright areas more uniform. This step is like clearing the fog from a window, allowing the faint outlines of the targets to become visible. Once the image is clearer, the computer looks for the start and end points of the line, which are often marked by the researchers or found automatically. From these points, the algorithm starts to trace the path. In older methods, the computer would simply look for the brightest neighboring pixel to decide where to go next. This works well when the image is perfect, but in the noisy underwater world, a random bright speck of noise could trick the computer into taking a wrong turn, causing the traced line to jump off course or break apart.

The new method adds a layer of intelligence by considering two types of direction. First, it looks at the immediate direction the line is taking at the very next step. If the line was moving straight, the computer expects the next step to be straight as well, and it gives a higher score to pixels that fit this expectation. Second, and perhaps more importantly, it looks at the direction over a slightly longer stretch. It asks whether the line, if extended a few steps forward, would still look like a straight segment. This "local approximate linear direction" acts as a stronger anchor. Even if a single step is obscured by noise or a gap in the image, the longer view helps the computer understand that the line is likely continuing in the same general direction. By combining the immediate direction with this longer-term trend, the system can bridge gaps in the image and ignore false alarms caused by random noise.

The researchers tested this approach on real sonar images of oil pipelines and submarine cables. In some cases, the targets were bright and clear; in others, they were extremely faint, broken into pieces, or hidden by the confusing reflections of the seabed. When they compared their new method against older techniques, the difference was clear. The traditional algorithms often struggled with the faint or broken lines, leaving gaps or following the wrong path entirely. The new method, however, successfully connected the broken pieces and traced the faint lines with high accuracy. It managed to find the true path of the pipeline even when the image quality was poor, demonstrating that using direction as a guide is a powerful way to cut through the confusion of underwater noise.

This work suggests that by respecting the natural geometry of the objects we are trying to find, we can see them more clearly even when the data is imperfect. The method does not require the image to be perfect; it works because it understands that a cable or pipe has a physical continuity that noise does not. The results show that this approach is robust and adaptable, capable of handling everything from strong, clear signals to weak, fragmented ones. While the researchers note that their current work focuses on two-dimensional images, they see a path forward to apply these same principles to three-dimensional sonar data, which could further improve our ability to map and monitor the complex world beneath the waves. The study confirms that a simple, logical understanding of how these lines move can lead to a much more reliable way of seeing them.

Technical Summary: Curve Target Extraction from Synthetic Aperture Sonar Image Based on Joint Direction Weighting

Problem Statement
Synthetic Aperture Sonar (SAS) is critical for underwater target segmentation and seabed analysis. However, SAS images often suffer from poor quality due to complex hydrographic conditions, resulting in low contrast, uneven brightness, blurred edges, and significant noise. These factors make the extraction of linear targets (e.g., cables and pipelines) challenging. Existing methods face specific limitations:

Edge detection and threshold segmentation often fail to achieve complete segmentation of linear targets, especially when gray values are low.
Hough Transform (HT) algorithms are primarily suited for straight lines; detecting curved targets requires complex segmented processing or projection mapping, which is sensitive to grayscale variations and prone to errors.
Active contour models (Snake, Level Sets) and graph cuts demand high image quality and exhibit high computational complexity, often failing when targets have voids, overlaps, or severe noise.
Minimal path and sequential tracking methods require high signal-to-noise ratios and clean data, struggling with low-gray-value and fractured targets.

There is an urgent need for a method that is simple, fast, and accurate, capable of extracting linear targets from SAS images even under conditions of low contrast, discontinuity, and severe noise.

Methodology
The authors propose a linear target extraction method based on the joint weighting of direction information. The approach consists of three main stages:

Preprocessing:
- Denoising: Total Variation (TV) denoising is applied to mitigate noise and reverberation while preserving edges and details.
- Brightness Equalization: A proportional mean brightness equalization method is used to flatten the background and enhance contrast, making weak targets more discernible.
Endpoint and Constraint Selection:
- Dual-threshold segmentation (based on Otsu's algorithm) is applied to the preprocessed image to generate candidate points.
- Users or the system select specific endpoints and a small number of constraint points to guide the extraction, ensuring the path passes through annotated locations.
Direction-Weighted Extraction Algorithm:
The core innovation lies in constructing an energy function that integrates three probabilistic components:
- Gray Value Probability: Pixels with higher gray values in the neighborhood are assigned higher probabilities of belonging to the target.
- Current Time Direction Weighting: Based on the assumption that real linear targets do not undergo abrupt directional changes between consecutive time steps, the algorithm penalizes large angular changes ( $\theta$ ) between the current step and the previous step. This is modeled using a cosine-based weighting function, $\omega(\theta) = \cos^n(\theta)$ .
- Local Approximate Linear Direction Weighting: To handle severe noise and fractures, the algorithm considers the curvature (approximated as slope) over a local path segment of $d$ pixels. It calculates the variation in slope ( $\beta$ ) and applies a weighting function, $\lambda(\eta) = \zeta^{-\beta}$ , where a small change in slope yields a high probability.
The final objective function combines these factors. The algorithm performs a recursive path search (top-down or bottom-up) to find the path that maximizes the accumulated energy, effectively connecting fragmented targets by extending the trajectory based on directional consistency rather than relying solely on pixel intensity.

Key Contributions

Joint Direction Weighting: The paper introduces a novel model that fuses current instantaneous direction changes with local approximate linear directions. This dual-constraint approach distinguishes the method from traditional grayscale-only searches or Hough-based methods.
Robustness to Fractures and Low Gray Values: By incorporating local slope continuity, the method can bridge gaps in fractured targets and extract low-gray-value objects that cause conventional algorithms to fail.
Efficiency: The method avoids the high computational complexity of active contour models and the sensitivity of Hough transforms to curvature fitting orders, offering a faster alternative suitable for large volumes of SAS data.

Experimental Results
The authors validated the method using real-world SAS datasets containing oil pipelines and submarine cables under varying conditions:

High/Low Gray-Scale Targets (No Discontinuities): The proposed method successfully extracted both high-energy and low-energy multipath reflections, outperforming comparison algorithms that showed errors in the right-side (low gray) targets.
Discontinuous Targets: In scenarios with overlapping and discontinuous targets, the proposed algorithm achieved complete and accurate extraction, whereas comparison algorithms struggled to connect discontinuities.
Severely Fragmented, Ultra-Low Gray-Scale Targets: In cases where targets were nearly invisible and highly fragmented, conventional algorithms failed completely. The proposed method successfully reconstructed the target paths by leveraging directional continuity, achieving relatively accurate extraction despite minor errors attributed to the extreme poor image quality.

Significance and Claims
The paper claims that the proposed method offers strong robustness and applicability for extracting linear targets from SAS images. It asserts that the method can:

Accurately extract linear targets regardless of whether they are straight or curved.
Handle challenging conditions including low signal-to-noise ratios, low contrast, and severe fragmentation.
Provide a reliable data source for subsequent research, analysis, and inversion processing of underwater linear targets.

The authors conclude that while the method is effective for 2D SAS images, future work will focus on extending this approach to three-dimensional SAS images to enable the extraction of arbitrary linear targets from any sonar image.

Technical Summary: Curve Target Extraction from Synthetic Aperture Sonar Image Based on Joint Direction Weighting

More like this