While camouflage gives wildlife and military vehicles a strategic survival advantage, it poses challenges, both for human and computer vision systems. For more on this topic, see our article on computer vision for disaster responses. It is difficult enough to detect objects designed to blend with their environments, but when the objects have blurry edges, the detection process is even more problematic.
However, in a paper written for the 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jinpeng He, Biyuan Liu, and Huaixin Chen of the University of Electronic Science and Technology of China propose a novel solution: Hourglass Vision Transformer with Dual-path Feature Pyramid, or HDPNet. He, Liu, and Chen’s research reveals that HDPNet outperforms 25 other methods, particularly for smaller objects and those with relatively indistinct boundaries.
Camouflaged object detection (COD) systems determine the boundaries of target objects, despite efforts by nature or humans to blur the lines of distinction.
Suppose a doctor is trying to detect a polyp during a medical examination. A COD system has to determine where the polyp ends and the lining of the colon begins. As another example, a military drone equipped with a COD system needs to identify a camouflaged tank even if it’s hidden by the flora of a jungle canopy.
Computer imaging presents other challenges because it may not render an object in sufficient detail for a traditional convolutional neural network (CNN) to distinguish it from its background. CNNs often focus on the most obvious features and may overlook some of the lower-level details.
For instance, a CNN may do a good job of identifying a tank and distinguishing it from a truck. However, it may not be able to tell a Russian tank from a Ukrainian one if the image lacks obvious markers.
As another COD solution, transformer-based methods perform well when they need to understand the global properties of a large image, getting a reliable “big picture” perspective. However, to gain a high-level understanding of an image, transformer-based methods have to divide the image into many smaller, low-resolution images. This can cause the loss of some important local details, such as differences in the armor plating of Russian and Ukrainian tanks.
HDPNet uses an hourglass architecture to both capture global semantic cues and extract detailed feature maps. The “hourglass” facet of the architecture describes how the Hourglass Vision Transformer (HVT) starts by capturing a wide swath of an image, then narrows down its analysis to focus on more granular details, and then widens out again by including the semantic details in context with the image in its entirety.
In effect, the system zooms out to gather enough information, then zooms in to identify the most important details, then zooms out again, providing both a comprehensive and extremely detailed analysis of the image.
To keep the vast amount of information the HVT encoder produces intact, HDPNet uses a Dual-Path Feature Pyramid Decoder (DPFD). This prevents important cues in an image from being diluted during analysis, resulting in a more accurate, richer image.
In addition, the HDPNet’s Feature Interaction Enhancement Module (FIEM) identifies the connections between the local details and the general camouflaged region. This ensures that detailed elements (such as texture or specific patterns in the image) and global features (such as the object’s overall shape) complement each other.
HDPNet systematically processes information across different scales while simultaneously maintaining the relationships between image details, regardless of the scale at which they were observed.
In this way, HDPNet can detect very small objects and distinguish minute details of larger, camouflaged items. This paves the way for doctors to distinguish symptoms of illness, scientists to identify camouflaged insects and other creatures, and military personnel to pinpoint assets of interest—all with greater accuracy.
Disclaimer: The authors are completely responsible for the content of this article. The opinions expressed are their own and do not represent IEEE's position nor that of the Computer Society nor its Leadership.
