SFE-YOLO: Spatial-Frequency Dual-Domain Enhanced YOLO for SAR Ship Detection
To address the limitations of existing detectors in complex marine environments, the paper proposes SFE-YOLO, a dual-domain enhanced YOLO framework that integrates pinwheel convolution for multi-orientation feature extraction and a frequency-domain block for sea clutter suppression, achieving superior performance in SAR ship detection on SSDD and HRSID datasets.
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
The ocean is a vast, shifting canvas where the line between a ship and a wave can blur into indistinguishable noise. For decades, monitoring this domain has relied on two main tools: cameras that see the world as we do, and radar that sees it through the eyes of physics. Cameras offer rich detail but fail when the sun sets or a thick fog rolls in. Radar, however, works day and night, piercing through clouds and darkness by bouncing radio waves off the surface. This makes it the only reliable eye for watching ships in the most treacherous weather. Yet, radar images are notoriously difficult to read. They are filled with static and grain, a visual chaos known as speckle noise, where the texture of a choppy sea can look exactly like the metal hull of a vessel. Distinguishing a tiny boat from a patch of rough water has long been a struggle for the computers tasked with scanning these images, often leading to missed detections or false alarms that could have serious consequences for safety and security.
Researchers have turned to artificial intelligence to solve this, teaching computers to recognize patterns in these grainy pictures. The most common approach uses a type of digital filter that scans an image in small squares, looking for edges and shapes. While this works well for many tasks, it struggles with the specific nature of ships. A ship is a long, thin object that can be pointing in any direction, but the standard filters are rigid and square, designed to catch round or blocky shapes. They often miss the slender lines of a ship's hull if it is tilted or if the image is too noisy. Furthermore, these standard filters only look at the immediate neighborhood of pixels, missing the bigger picture of the entire scene, which makes it hard to tell the difference between a real ship and a confusing patch of sea clutter.
To address these specific weaknesses, a team of researchers from institutions in China has developed a new detection system called SFE-YOLO. Instead of relying on a single way of looking at the image, they created a dual-domain framework that examines the picture in two different ways simultaneously. The first part of their system tackles the shape problem. They replaced the standard square filters with a new type of filter that acts like a pinwheel, capable of scanning in multiple directions at once. This allows the computer to trace the long, thin contours of a ship regardless of which way it is facing, capturing details that the old, rigid filters would simply skip over.
The second part of the system addresses the noise problem by changing how the computer processes the image's information. Instead of just looking at the pixels as they appear, the system temporarily translates the image into a different language based on frequency, which is a way of describing how quickly the brightness of the image changes from spot to spot. In this frequency view, the steady, repetitive patterns of sea noise stand out clearly from the sharp, unique edges of a ship. The researchers built a module that can identify and dampen the low-frequency noise of the ocean while simultaneously boosting the high-frequency signals that define the ship's outline. After cleaning up the image in this frequency view, the system translates it back to the normal picture, now much clearer and easier to read.
When the team tested this new method on two major collections of radar ship images, the results were significant. On a standard dataset containing over a thousand images, their system correctly identified ships with a high degree of accuracy, achieving a score of 0.973 on a scale where 1.0 is perfect. On a more difficult, high-resolution dataset with nearly six thousand images, which includes many tiny ships and complex harbor backgrounds, the system maintained a strong performance with a score of 0.885. More importantly, the system proved particularly good at finding small vessels, a task where previous models often failed. It reduced the number of times it missed a small boat and also cut down on the number of times it mistakenly identified a wave as a ship.
The researchers found that combining these two approaches—the directional scanning for shape and the frequency-based cleaning for noise—created a synergy that was greater than the sum of its parts. The system did not just get slightly better; it became robust enough to handle the chaotic reality of the open sea. By teaching the computer to see the ship's structure from every angle and to ignore the background chatter of the ocean, the new method offers a more reliable way to monitor maritime traffic. This advancement suggests that in the future, automated systems could provide more consistent and accurate surveillance for law enforcement and search-and-rescue operations, ensuring that even the smallest vessels are not lost in the noise of the sea.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.