💻 computer science

GHD-DETR: Gated Hierarchical Dynamic Architecture for Object Detection Under Complex Imaging Conditions

The paper proposes GHD-DETR, a novel detection transformer architecture featuring GatedNestStage, BDB, and HAGF modules to significantly enhance object detection robustness for small, occluded, and low-contrast objects under complex imaging conditions, achieving substantial performance improvements on challenging datasets like RTTS, HazyDet, and DUO.

Original authors: Zidian Wei, Wei Feng, Linghan Jiang, Zunwang Ke

Published 2026-09-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zidian Wei, Wei Feng, Linghan Jiang, Zunwang Ke

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

Autonomous vehicles rely on a pair of digital eyes to navigate the world, constantly scanning the road to identify cars, pedestrians, and cyclists. For these systems to function safely, they must recognize objects with absolute precision, even when the view is imperfect. In the real world, however, the camera's view is often compromised by fog, heavy rain, or murky water, which scatter light and wash out colors. When an image becomes hazy or blurry, the subtle edges that define a vehicle or a person begin to dissolve, making it difficult for standard computer vision software to tell where one object ends and the background begins. This is not merely a matter of poor picture quality; it is a critical safety failure. If a self-driving car cannot distinguish a pedestrian from a fog bank, the consequences can be catastrophic. Researchers have long sought to build detection systems that do not just see better, but understand better, even when the visual information is degraded and incomplete.

A team of researchers from Xinjiang University has developed a new system designed specifically to solve this problem of seeing clearly through the blur. They call their creation GHD-DETR, a sophisticated architecture that acts as a robust filter for visual data, allowing a computer to detect objects in conditions where traditional methods often fail. The system was tested against three distinct and challenging environments: real-world foggy traffic scenes, dense urban fog captured by drones, and murky underwater footage where visibility is severely limited. In every case, the new system outperformed existing methods, demonstrating a significant ability to recover the shapes and positions of objects that were previously lost to the haze.

The core of this success lies in how the system processes information. Standard detection tools often treat all parts of an image equally, which causes them to get confused when noise and blur are present. The new approach, however, uses a specialized structure that acts like a multi-layered sieve. It first breaks the image down into different sizes to catch both large shapes and tiny details. Within this structure, a component called the GatedNestStage acts as a gatekeeper. It examines the incoming visual data and decides which parts are useful and which are just noise caused by the weather. By filtering out the visual static while preserving the critical edges of an object, it ensures that the system focuses only on the information that matters for identification.

Once the system has isolated the useful features, it must combine them to form a complete picture of the scene. This is where the second innovation comes into play. The researchers introduced a module that splits the processing into two parallel paths. One path looks at the big picture to understand the general context, while the other path zooms in to capture the fine, local details that are often lost in fog. These two streams of information are then merged together in a way that ensures the system does not lose the sharpness of the object's boundary while still understanding its overall shape. This dual-path approach prevents the system from becoming confused by the low contrast and blurring that typically plague foggy or underwater images.

The final piece of the puzzle is a mechanism that decides how to blend these different layers of information together. In many systems, simply stacking layers of data can lead to a muddy result where important details are drowned out. The new system uses a dynamic fusion method that weighs the importance of each piece of information as it is combined. It acts like a skilled editor, knowing exactly when to emphasize the high-level shape of a car and when to highlight the specific texture of a pedestrian's clothing. This ensures that the final output is a clear, confident identification of the object, regardless of how degraded the original image was.

The researchers tested their system on three difficult datasets to prove its worth. The first set consisted of over 4,000 real-world images of traffic in heavy fog, where visibility is often reduced to a few meters. The second set included thousands of drone-captured images of cities shrouded in dense haze, featuring small vehicles that are particularly hard to spot. The third set involved underwater images, where the water itself distorts colors and obscures boundaries. In the foggy traffic test, the new system improved the accuracy of detecting cars by more than 12 percentage points compared to the previous best method. For buses, the improvement was even more dramatic, jumping by nearly 30 percentage points. In the dense fog scenarios, the system showed an overall improvement of over 9 percentage points, and in the underwater tests, it still managed to improve accuracy by more than 4 percentage points.

These results indicate that the system is not just a minor tweak but a fundamental shift in how machines can perceive the world under stress. By explicitly designing the architecture to handle the specific ways that fog and water distort light, the researchers have created a tool that is far more resilient than its predecessors. The system does not try to clear the fog or clean the water; instead, it learns to see through the distortion by focusing on the structural integrity of the objects within the scene. This approach suggests that the future of autonomous driving in adverse weather may not depend on better cameras, but on smarter ways of interpreting the imperfect images those cameras capture. The work confirms that with the right design, machines can maintain their situational awareness even when the world around them becomes difficult to see.

Technical Summary: GHD-DETR

Problem Statement

Computer vision systems for autonomous driving face significant performance degradation under complex imaging conditions, particularly in adverse weather such as fog and underwater environments. These conditions introduce challenges including reduced visibility, low contrast, image blur, and heavy occlusion. Conventional object detection algorithms often struggle to accurately capture target features in these scenarios, leading to false positives, missed detections, and imprecise localization of critical objects like vehicles and pedestrians. Existing Transformer-based detectors (e.g., DETR variants) and CNN-based approaches often suffer from limitations such as quadratic computational complexity, static feature aggregation that fails to adapt to scene degradation, and a lack of dynamic modulation to suppress noise while preserving structural details.

Methodology

The authors propose GHD-DETR (Gated Hierarchical Dynamic Detection Transformer), a novel architecture designed to enhance detection robustness in degraded imaging conditions. The framework integrates three primary optimized modules:

GatedNestStage (Backbone):
- Replaces standard stage designs with a nested, dual-gated block architecture.
- Employs Adaptive Gating using depthwise convolutions and activation-guided Hadamard products to modulate channel-spatial information dynamically.
- Utilizes a Deep Transformation stage with re-parameterized convolutions, channel attention, and gated normalization within a residual pipeline.
- Goal: To robustly extract multi-scale semantic features, suppress noise, and enhance structural details specifically under low-visibility conditions.
Bifurcated Dynamic Block (BDB):
- An adaptive encoding module designed to replace inefficient global attention mechanisms (like standard self-attention) which suffer from $O(N^2)$ complexity.
- Global Interaction & Local Refinement: Uses a Linear Attention Module (LAM) with explicit kernel mapping to achieve linear complexity $O(N)$ while capturing global dependencies. It includes Dynamic Tanh Normalization to model contrast degradation.
- Multi-scale Convolution Adapter: Employs parallel depthwise convolutions (3×3, 5×5, 7×7) to capture multi-scale spatial context.
- Feature Mixing & Secondary Refinement: Performs spatial-channel collaborative enhancement.
- Goal: To enhance feature discriminability under occlusion, blur, and low contrast while maintaining computational efficiency.
Hierarchical Attention Gated Fusion (HAGF):
- A multi-scale fusion module that rewires the standard Feature Pyramid Network (FPN) pathway.
- Integrates dual local-global attention mechanisms to extract enhanced features with spatial awareness.
- Utilizes learnable gating weights generated via global average pooling to adaptively select and enhance features.
- Incorporates a detail-enhanced convolution with residual connections to preserve fine-grained spatial cues.
- Goal: To preserve spatial details while reinforcing high-level semantic representations, preventing the mutual interference between low-level noise and high-level semantics during fusion.

Key Contributions

The paper outlines three main contributions:

GatedNestStage: A robust feature extraction architecture employing dual-gated modulation and nested multi-scale design. It enables dynamic feature calibration and adaptive fusion, significantly improving performance in hazy conditions by suppressing noise and enhancing structural details.
Bifurcated Dynamic Block (BDB): An efficient adaptive encoding module using bifurcated dynamic aggregation and gated multi-branch modulation. It simultaneously enhances semantic representation and spatial-channel structural preservation, improving robustness in occluded and low-contrast scenarios.
Hierarchical Attention Gated Fusion (HAGF): A multi-scale fusion module employing hierarchical attention gating and dynamic feature selection. It significantly enhances feature fusion efficiency and multi-scale object perception in complex scenes.

Experimental Results

The authors evaluated GHD-DETR on three challenging datasets: RTTS (real-world hazy images), HazyDet (dense fog scenes with UAV captures), and DUO (underwater environments).

RTTS Dataset: The proposed method achieved an Average Precision (AP) of 0.475 and AP50 of 0.716. This represents an improvement of 3.1 percentage points in AP over the baseline D-FINE (0.444) and outperformed other state-of-the-art methods like FA-YOLO (AP50 59.7% vs. 71.6%). Notable gains were observed in detecting cars, motorbikes, and buses.
HazyDet Dataset: GHD-DETR achieved an AP of 0.467 and AP50 of 0.645, surpassing D-FINE by 3.9 percentage points in AP and 4.5 percentage points in AP50.
DUO Dataset: The model achieved an AP of 0.643 and AP50 of 0.847, improving upon D-FINE by 2.6 percentage points in AP and 3.7 percentage points in AP50.
Ablation Studies: Experiments confirmed that each module contributes positively. The BDB module provided the largest single gain (1.6% AP improvement over baseline), while the combination of all three modules (A+B+C) yielded the best overall performance, demonstrating complementary effects.
Visualization: Heatmap visualizations indicated that GHD-DETR produces sharper boundaries and more focused attention responses compared to baseline models (e.g., Deformable DETR, DINO, YOLOX), particularly in dense fog and turbid water where other methods exhibited scattered or drifting attention.

Significance and Claims

The paper claims that GHD-DETR effectively addresses the fundamental challenges of detecting small-scale, heavily occluded, and low-contrast objects in degraded imaging scenarios. By integrating gated hierarchical mechanisms, bifurcated dynamic encoding, and attention-gated fusion, the architecture achieves a superior balance between detail preservation and semantic enhancement.

The authors assert that the proposed method demonstrates significant efficacy in addressing diverse image quality degradation, validating its potential for real-world deployment in complex autonomous driving environments where visibility is compromised by fog, haze, or underwater conditions. The work is presented as a step forward in making computer vision systems more reliable under adverse environmental conditions, with code and models made publicly available to facilitate further research.