💻 computer science

DDS-YOLO: A Detail-Enhanced and Dynamic Scale-Adaptive Network for UAV Small Object Detection

This paper proposes DDS-YOLO, a dynamic scale-adaptive network featuring detail enhancement, adaptive feature fusion, and a lightweight high-resolution branch to significantly improve the accuracy and efficiency of small object detection in UAV imagery.

Original authors: Qiang Liu, Chen Yuanze, Niu Dexue

Published 2026-10-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qiang Liu, Chen Yuanze, Niu Dexue

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

From the vantage point of a drone, the world below transforms into a shifting mosaic of tiny details. A car becomes a speck, a pedestrian a blur, and a distant building a mere patch of color. For computers tasked with seeing these scenes, this perspective presents a unique and stubborn challenge. Objects that are large and clear to the human eye shrink into a handful of pixels for a machine, losing the edges and textures that define them. When these small targets are crowded together or hidden behind complex backgrounds, standard vision systems often fail to distinguish them from the noise. This limitation hinders the ability of unmanned aircraft to perform critical tasks like monitoring traffic, surveying disaster zones, or inspecting infrastructure, where missing a small detail can have significant consequences.

To solve this, researchers have long relied on deep learning, a method where computers learn to recognize patterns by studying thousands of examples. Among the tools used for this, a family of systems known as YOLO has become a favorite for its speed and balance between accuracy and efficiency. However, even these advanced systems struggle when the objects they are looking for are extremely small and vary wildly in size. A new study by Qiang Liu and colleagues at the Shandong University of Aeronautics addresses this specific gap. They have developed a refined version of the system, named DDS-YOLO, designed specifically to keep the fine details of tiny objects visible while adapting to the changing scales found in aerial photography.

The core problem the researchers tackled is that as a computer analyzes an image, it naturally simplifies the picture to understand the big picture. In doing so, it often discards the very information needed to spot a small object: its sharp boundaries and subtle textures. The team's solution involves three distinct strategies working together to preserve and enhance this lost information. First, they introduced a method to sharpen the image's details before the computer even begins its main analysis. This process acts like a filter that isolates the edges of objects and the high-frequency textures that define them, while simultaneously ignoring the distracting background noise. By focusing on these specific cues, the system ensures that the faint outline of a distant person or a small vehicle is not washed out during the initial stages of processing.

Next, the researchers addressed the issue of scale. In a single drone image, a large truck might appear close and clear, while a motorcycle far away looks like a tiny dot. Traditional systems often treat these different sizes with a rigid, one-size-fits-all approach, which can lead to missed detections. The new system, however, learns to adjust its focus dynamically. It analyzes the image to determine which parts of the scene contain small objects and which contain larger ones, then weights the information accordingly. This allows the system to give more attention to the high-resolution details needed for tiny targets without losing the broader context required for larger ones. It is a flexible approach that changes its strategy based on what is actually present in the scene, rather than following a fixed set of rules.

Finally, to catch the smallest objects that might otherwise slip through the cracks, the team added a dedicated pathway for high-resolution prediction. Standard systems often downsample images too early, making it impossible to see objects that occupy only a few pixels. The new design keeps a separate, high-resolution stream of information alive throughout the process, specifically tuned to detect these extremely small targets. This addition works in tandem with a more precise method for drawing the boxes around detected objects, ensuring that even the most elusive targets are located with greater accuracy.

The results of this work were tested on a large collection of real-world drone images containing thousands of diverse scenes, from busy city streets to open fields. The new system significantly outperformed the previous standard, improving its ability to find small objects by nearly ten percentage points in terms of precision and recall. In practical terms, this means the system is far less likely to miss a pedestrian or a vehicle in a crowded scene. It achieved these gains while maintaining a speed of 161 frames per second, which is fast enough for real-time applications. The model remains compact, requiring less computing power than many of its competitors, making it suitable for deployment on the limited hardware found on actual drones.

The study confirms that by combining detail enhancement, dynamic scale adjustment, and a specialized high-resolution branch, it is possible to overcome the inherent difficulties of aerial small-object detection. The researchers did not rely on massive increases in computing power to achieve this; instead, they refined how the system processes information. The findings suggest that for unmanned aerial vehicles to become truly effective in complex environments, they must be able to see the small things clearly. This new approach provides a clear path forward, offering a system that is both fast and remarkably sensitive to the tiny details that matter most from the sky.

Technical Summary: DDS-YOLO for UAV Small Object Detection

1. Problem Statement

Unmanned Aerial Vehicle (UAV) imagery presents unique challenges for object detection, particularly for small objects. These challenges stem from:

Limited Pixel Representation: Small objects occupy few pixels, resulting in weak texture, boundary, and structural cues that are easily degraded during deep feature extraction.
Severe Scale Variation: Changes in flight altitude, viewing angles, and object distances cause significant intra-class scale variations, making consistent detection across scales difficult.
Complex Backgrounds and Density: Dense distributions, occlusion, and complex backgrounds reduce foreground-background contrast, leading to false and missed detections.
Resource Constraints: UAV platforms have limited onboard computation and energy, necessitating models that balance high accuracy with strict efficiency requirements.

Existing methods often address these issues in isolation. Feature enhancement methods may struggle when source details are scarce, while multi-scale fusion approaches often rely on fixed paths that cannot adapt to changing object distributions. Furthermore, many solutions fail to jointly preserve fine details, adapt to scale variations, and maintain real-time inference speeds within a compact framework.

2. Methodology

The authors propose DDS-YOLO, a network built upon the YOLOv11n baseline. The architecture integrates three primary strategies to address detail loss, scale variation, and weak responses for extremely small objects, alongside an optimized loss function.

2.1 Detail Enhancement Strategy (DES)

Placed at the transition from shallow to intermediate features, DES aims to preserve boundaries and textures while suppressing background noise. It operates through three parallel branches:

Local Structure Branch: Uses a $3\times3$ depthwise convolution to model local boundaries and textures.
Contextual Branch: Employs a dilated $3\times3$ convolution (dilation rate 2) to enlarge the receptive field and incorporate surrounding context.
High-Frequency Residual Branch: Extracts structural residuals by subtracting a local average-pooled response from the input, capturing rapidly varying components like corners and edges.

These branches are adaptively weighted using a channel-wise attention mechanism (Global Average Pooling followed by a Softmax mapping) to generate input-dependent weights ( $\alpha_l, \alpha_c, \alpha_h$ ). A spatial selection map is then generated to suppress irrelevant background regions before fusing the enhanced features with the original feature flow via a residual connection.

2.2 Dynamic Scale Adaptation Strategy (DSAS)

Replacing fixed top-down and bottom-up fusion nodes in the neck, DSAS dynamically aligns and fuses adjacent feature levels ( $F_{l-1}, F_l, F_{l+1}$ ) based on the current input distribution.

Alignment: Features are aligned to a common spatial resolution and channel dimension using $1\times1$ convolutions, downsampling, and upsampling.
Dynamic Weighting: A global descriptor is extracted from the concatenated aligned features. A lightweight network generates channel-wise weights ( $\beta$ ) for each scale branch via Softmax, allowing the model to emphasize high-resolution details, current-level information, or low-resolution semantics as needed.
Local Correction: A spatial response map is generated to correct regional scale imbalances, applied via a residual addition to the weighted fusion.

2.3 Small Object Optimization Strategy (SOOS)

To address the loss of extremely small objects (which may occupy only one cell in the standard P3 layer after downsampling), SOOS introduces a lightweight P2 prediction branch (stride 4).

Architecture: It combines the shallow high-resolution P2 feature from the backbone with the upsampled P3 semantic feature from the neck.
Fusion: A "LiteFusion" unit (combining depthwise and pointwise convolutions) merges these features to generate the P2 prediction feature.
Head: A lightweight decoupled detection head with separate classification and regression paths uses depthwise separable convolutions to manage the computational cost of high-resolution processing.

2.4 Loss Function

The model retains the YOLOv11 classification and Distribution Focal Loss but replaces the bounding-box regression loss with WIoU v3 (Wise Intersection over Union version 3). WIoU v3 utilizes dynamic non-monotonic focusing to reduce the influence of low-quality samples and emphasize samples with high regression value, improving localization without adding inference-time parameters.

3. Key Contributions

The paper outlines three main contributions:

DES: A module that strengthens local structures, context, and high-frequency cues to preserve small-object boundaries and textures while adaptively suppressing background responses.
DSAS: A mechanism that dynamically adjusts the contributions of adjacent feature levels to accommodate changing object-scale distributions, replacing fixed fusion paths.
SOOS: The introduction of a lightweight P2 branch to improve the prediction of extremely small objects, coupled with WIoU v3 for enhanced bounding-box regression.

4. Experimental Results

Experiments were conducted on the VisDrone2019-DET dataset (6,471 training images, 548 validation images). The model was compared against various lightweight YOLO variants and state-of-the-art UAV detectors.

Performance Metrics

DDS-YOLO achieved the following results on the VisDrone2019 validation set:

Precision: 56.0% (Improvement of +9.8% over YOLOv11n)
Recall: 44.5% (Improvement of +9.4% over YOLOv11n)
mAP0.5: 46.3% (Improvement of +8.9% over YOLOv11n)
mAP0.95: 28.6% (Improvement of +6.7% over YOLOv11n)
Small Object AP (APS): 15.4%

Efficiency

Despite the added modules, the model maintains a favorable accuracy-efficiency trade-off:

Parameters: 8.7 M
FLOPs: 23.8 G
Inference Speed: 161 FPS (on an NVIDIA RTX 4080)

Comparisons with other methods (e.g., WCDB-YOLO, EdgeYOLO-S, DFA-DETR) show that DDS-YOLO achieves higher accuracy with significantly fewer parameters and lower computational costs, particularly for small objects.

5. Significance and Claims

The paper claims that DDS-YOLO provides a favorable accuracy–efficiency trade-off specifically tailored for real-time UAV small object detection. The significance of the work lies in its ability to:

Simultaneously address detail preservation, scale adaptation, and background suppression within a single compact framework.
Improve detection stability for distant, weak, and densely distributed small objects without relying on large-scale model expansion.
Maintain real-time performance (161 FPS) suitable for onboard UAV deployment, outperforming heavier Transformer-based detectors (like RT-DETR) and other specialized YOLO variants in both accuracy and speed.

Visual analysis confirms that the model recovers more small vehicles and pedestrians in complex scenes compared to the baseline, validating the complementary roles of the proposed strategies. The authors assert that the improvements are driven by the specific design of the DES, DSAS, and SOOS modules, as confirmed by ablation studies.