Related work

D-FINE Framework

D-FINE (Dynamic Fine-Grained Matching for End-to-End Object Detection) consists of two key components: Fine-Grained Distribution Refinement (FDR) and Global Optimal Localization Self-Distillation (GO-LSD). The core idea of D-FINE is to alleviate the sparse supervision and slow convergence caused by the traditional one-to-one matching strategy in DETR during the early stages of training, while preserving the original end-to-end detection framework. Specifically, D-FINE no longer restricts supervision to a single optimal match. Instead, it dynamically introduces multi-scale and multi-candidate high-quality matching relationships during training according to prediction quality, thereby providing denser and more stable gradient feedback for decoder queries and significantly improving both convergence speed and detection accuracy.

At present, extensions of D-FINE have mainly focused on specific application scenarios, and this line of research is still at an early stage. SAR-D-FINE [27] is designed for detecting small and densely distributed ships in synthetic aperture radar (SAR) imagery. Built upon the D-FINE framework, it introduces a context-aware feature extraction strategy and enhances multi-scale feature fusion, thereby improving the capture of contextual information and the representation of small objects. As a result, the model achieves favorable performance in dense ship detection tasks. However, this method still relies heavily on the characteristics of SAR imagery and the distribution of specific datasets, which may limit its generalization capability in more complex or diverse remote sensing scenarios. IW-D-FINE [28] is developed for ocean internal wave detection in SAR images. Based on the D-FINE architecture, it adopts an anchor-free one-stage detection framework and incorporates several improved modules, including the SimAM attention mechanism, the MultiScalePCA feature fusion module, and a MetaFormer-based encoder structure. These components enhance multi-scale feature extraction and achieve a better balance between local and global feature modeling. Experimental results show that IW-D-FINE attains high detection accuracy while maintaining real-time inference efficiency. HF-D-FINE [29] is a D-FINE-based method for tiny-object detection in UAV imagery. It enhances high-resolution feature representation through the HF Hybrid Encoder, performs cross-scale feature fusion using the CAF module, and introduces the Outer-SNWD loss to improve bounding box regression accuracy for tiny objects. These designs significantly improve detection performance on UAV datasets. However, the method still increases computational complexity to some extent, and its performance gains depend on the trade-off between richer high-resolution feature representations and inference efficiency.

Motivated by these limitations, particularly the suboptimal balance between efficiency and accuracy in high-resolution tiny-object detection, the limited adaptability of existing D-FINE variants across diverse scenarios, and the need for more effective query optimization and multi-scale feature aggregation, we propose AQF-Net. Built upon the D-FINE framework, AQF-Net incorporates a Fixed-Query Self-Attention (FQSA) mechanism and a Large-Receptive-Field Enhancement (LREA) module to improve computational efficiency and multi-scale feature fusion under high-resolution settings. Furthermore, by integrating an adaptive query modeling strategy, AQF-Net optimizes query generation and target matching, thereby improving the detection accuracy and robustness of tiny objects.

Fixed-Query Self-Attention (FQSA)

Self-attention has shown remarkable capability in modeling long-range dependencies. However, its computational complexity typically grows quadratically with the spatial resolution of input features, resulting in substantial computational overhead in high-resolution dense prediction tasks. Although existing studies have attempted to alleviate this issue through strategies such as window partitioning and sparse connections, these methods often suffer from limited global context modeling or insufficient cross-region information interaction. To address these limitations, we propose a Fixed-Query Self-Attention (FQSA) mechanism. The core idea of FQSA is to generate queries in a fixed low-resolution space and perform attention aggregation with multi-scale keys and values, thereby decoupling the computational complexity of attention from the input resolution and effectively reducing computational cost while preserving a global receptive field.

The overall architecture of FQSA is illustrated in Fig 2. In this work, the input features are downsampled to a low-resolution feature map with a fixed spatial size of using Adaptive Average Pooling. Then, is fed into two parallel branches. In the Q branch, is linearly transformed to generate the query matrix, and local features are selectively enhanced through lightweight convolution. In the K&V branch, is processed through a pyramidal structure to extract multi-scale contextual features, and keys and values are generated via linear transformation. The resulting , , and are used to compute multi-head scaled dot-product attention in the low-resolution space. Finally, the attention outputs from multiple heads are fused through linear projection and channel alignment, and then restored to the original spatial resolution via bilinear interpolation upsampling, producing the final output feature .

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Architecture of FQSA.

https://doi.org/10.1371/journal.pone.0355299.g002

To overcome the computational bottleneck of standard self-attention under high-resolution settings, FQSA redesigns the generation of Q, K, and V by constraining the attention computation to a fixed-size low-resolution query space and a multi-scale key-value space. On this basis, FQSA still adopts multi-head scaled dot-product attention for feature aggregation, while all attention operations are performed in the low-resolution space:

(1)

Where, , , and are obtained from the low-resolution feature map through linear transformation. In FQSA, the sequence length of the query is fixed at , while the sequence lengths of keys and scale linearly with the input resolution. This design reduces the size of the attention matrix from in standard self-attention to , thereby reducing memory consumption and improving computational efficiency.

The computational complexity of FQSA can be decomposed into several components. Specifically, the computational complexity of the adaptive pooling and upsampling operations is denoted as , the complexity of the linear transformations as , and the complexity of the self-attention computation as . Therefore, the overall computational complexity of FQSA can be expressed as . Since denotes the side length of the low-resolution query space and is fixed as a constant (e.g., ), independent of the input sequence length , and is much larger than , the overall computational complexity can be simplified to . Compared with the computational complexity of standard self-attention, FQSA significantly reduces computational cost, enabling the model to efficiently handle high-resolution inputs.

In summary, by combining fixed-budget queries with multi-scale contextual keys and values, FQSA enables efficient attention computation within a compact representation space. This design not only preserves the ability to model global dependencies, but also effectively alleviates the excessive computational complexity of attention under high-resolution settings, thereby providing a practical solution for applying vision Transformers to dense prediction tasks.

Large-Receptive-Field Enhancement (LREA)

In UAV tiny-object detection, it is crucial to efficiently capture long-range contextual dependencies while preserving sensitivity to fine-grained features. Although existing architectures have made progress in long-range dependency modeling and receptive field expansion, several challenges remain. First, Vision Transformers [30] usually incur high computational cost, making them difficult to apply to efficient real-time processing. Second, efficient attention mechanisms such as SeaFormer [31] have reduced computational overhead to some extent, but they still face limitations in further performance optimization. In addition, large-kernel attention methods, such as SLaK [32] and LSKA [33], remain insufficient in terms of multi-scale feature fusion and dynamic receptive field adaptation. To address these issues, we propose a Large-Receptive-Field Enhancement (LREA) module, as illustrated in Fig 3. The design of LREA follows the mainstream attention-feedforward paradigm and mainly consists of two components: a Large Kernel Attention (LKA) mechanism and a Convolutional Feed-Forward Network (CFFN) [34]. Through the synergy of these two components, LREA expands the receptive field and captures global semantic information while effectively preserving fine details and boundary information, thereby improving both model accuracy and efficiency in UAV tiny-object detection.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Architecture of LREA.

https://doi.org/10.1371/journal.pone.0355299.g003

LKA innovatively combines convolutional operations with the attention mechanism and expands the receptive field through the proposed Sparse Decomposed Broad Convolution (SDBConv) module. To further improve the adaptability of feature fusion, we design a Bidirectional Selective Aggregation (BSA) module, which enables the network to dynamically adjust the fusion of features at different scales according to the input content. The Convolutional Feed-Forward Network (CFFN) is responsible for further refining and integrating features, thereby ensuring a balanced output representation with sufficient information preservation. Specifically, CFFN adopts pointwise convolution (PWConv) to expand and compress the channel dimension, while a depthwise convolution (DWConv) is embedded in the intermediate stage to enhance spatial information interaction.

Sparse Decomposed Broad Conv (SDBConv)

The SDBConv module is a core component of LREA. Through a sparse decomposition strategy, it decomposes a conventional large-kernel convolution into smaller convolutional units and two dilated convolution strips, thereby achieving an equivalent expansion of the receptive field.

As illustrated in Fig 4, the SDBConv module first employs a 5 × 5 depthwise separable convolution to extract local features and generate a small-kernel attention feature map . The computational cost of this operation is substantially lower than that of directly using an extremely large convolution kernel. Subsequently, two strip-shaped dilated convolutions are applied in parallel to further process this feature map : one adopts a 1 × 𝑘 horizontal convolution, while the other uses a 𝑘 × 1 vertical convolution. This design enables the module to capture broader contextual information and expand the receptive field along the spatial dimension, thereby effectively enhancing long-range dependency modeling. The outputs of these two branches are denoted as the horizontally extended attention feature map and the vertically extended attention feature map , respectively. These three complementary attention feature maps are then fed into the kernel selection mechanism for adaptive fusion, producing a fused attention feature map . To further strengthen the representational capability of the fused feature, a pointwise convolution is finally applied to perform feature transformation and generate the modulation weight map. Finally, the generated weight map is multiplied element-wise with the original input to accomplish the final attention modulation. Overall, by virtue of this decomposition strategy, SDBConv significantly enlarges the receptive field while maintaining high computational efficiency, enabling effective long-range dependency modeling and providing efficient support for the large-kernel attention mechanism in LREA.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Architecture of SDBConv.

https://doi.org/10.1371/journal.pone.0355299.g004

Bidirectional Selective Aggregation (BSA)

The three attention feature maps generated by SDBConv emphasize information at different scales and orientations, respectively. To effectively integrate these complementary features, we propose a Bidirectional Selective Aggregation (BSA) module. Within LREA, the BSA module dynamically adjusts feature weights along both the channel and spatial dimensions, thereby significantly enhancing the model’s feature fusion capability. The overall structure of BSA is illustrated in Fig 5.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Architecture of BSA.

https://doi.org/10.1371/journal.pone.0355299.g005

Channel-weight branch: First, the three attention feature maps are summed to construct a shared attention feature :

(2)

Then, is processed by global average pooling, followed by lightweight pointwise convolution and depthwise convolution for feature mapping:

(3)

Finally, three independent expansion mappings are applied to to generate three groups of channel weights , and .

Spatial-weight branch: Average pooling and max pooling are applied to the shared attention feature along the channel dimension, and the resulting features are then concatenated:

(4)

Then, the feature map is fused and transformed through a 7 × 7 convolution to generate a spatial weight map:

(5)

is divided into three groups, denoted as , , and . Finally, the channel weights, spatial weights, and the original attention maps are combined through element-wise multiplication and addition:

(6)

The BSA module jointly models channel-wise and spatial-wise weights, enabling the network to dynamically adjust its receptive-field preference according to the input content. When target structures exhibit stronger horizontal continuity, and are enhanced; conversely, when richer vertical contextual information is required, and are enhanced. This mechanism allows for more expressive and adaptive multi-scale attention fusion.

LREA is particularly suitable for DETR-style detectors because the quality of encoder features directly affects query selection and subsequent bipartite matching. In UAV tiny-object detection, small targets are often weakly represented and easily confused with background textures. If the encoder features lack sufficient contextual discrimination, decoder queries may compete for ambiguous regions or be matched to redundant background responses. By expanding the receptive field through SDBConv and adaptively selecting contextual branches through BSA, LREA enhances object-relevant regions while preserving fine-grained spatial details. This improves the separability of tiny objects in the encoder memory and provides more reliable features for query initialization and matching in the decoder.

Adaptive query modeling strategy

In the decoding stage, the D-FINE Transformer still adopts a fixed upper bound K on the number of queries for query selection and decoding. This design exhibits notable limitations in scenarios where the number of objects varies significantly across images. In sparse scenes, a fixed number of queries may introduce redundant predictions, leading to increased false positives and unnecessary computational overhead. In contrast, in dense scenes, the detection recall can be constrained by the upper bound K. To address these issues in a systematic manner, we introduce an adaptive query modeling strategy within the AQF-Net framework. Inspired by the concept of Dynamic Query Selection [26], this strategy leverages predicted object count information to adaptively determine the number of queries required by the decoder, thereby enabling dynamic alignment between query allocation and scene complexity.

The overall architecture of the adaptive query modeling strategy is illustrated in Fig 6, which consists of three key components: the Categorical Counting Module (CCM), the Counting-Guided Feature Enhancement (CGFE) module, and the Dynamic Query Selection module. In particular, the PKIBlock from PKINet [35] is incorporated into the CCM to enhance multi-scale feature extraction and improve the robustness of object counting. The processing pipeline is as follows. First, the input features are processed through Projection and Flatten operations and then fed into the dynamic query mechanism. The CCM predicts the number of objects in the image and uses this information to estimate the number of queries required by the decoder. The predicted count is subsequently mapped to a specific query number through the dynamic query mapping process, which determines the actual number of queries used during decoding. Meanwhile, the CGFE module performs multi-scale weighted fusion on the density map generated by the CCM to further enhance feature representation. Finally, by dynamically adjusting the number of queries, the model adaptively allocates computational resources across different scenarios, thereby improving both detection performance and efficiency.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Architecture of adaptive query modeling strategy.

https://doi.org/10.1371/journal.pone.0355299.g006

Categorical counting module (CCM)

Within the adaptive query modeling strategy, the Categorical Counting Module (CCM) aims to predict the number of object instances in the input image and categorize them into predefined density levels, which are then used to guide dynamic query allocation. To enhance the representational capability of counting features while maintaining computational efficiency, a lightweight design is adopted for CCM by introducing efficient channel encoding and a multi-kernel convolution module, namely PKIBlock.

As illustrated in Fig 7, PKIBlock consists of two key submodules: a multi-scale depthwise convolution aggregation module and a Context Anchor Attention (CAA) module. The former captures local contextual information at multiple scales, while the latter incorporates global contextual cues to further enhance feature representation, thereby improving the accuracy and robustness of object counting.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Architecture of PKIBlock.

https://doi.org/10.1371/journal.pone.0355299.g007

The detailed computation process is as follows. Let the input feature be denoted as . The multi-scale depthwise convolution aggregation module adopts an Inception-style multi-branch structure, where depthwise separable convolutions with different kernel sizes are applied in parallel to capture local contextual information at multiple scales. Specifically, kernel sizes are selected, denoted as . For each kernel size , the following operation is performed:

(7)

Since each convolution operates independently on each input channel, this design significantly reduces the number of parameters. The outputs from all multi-scale branches are then fused with the original input feature via element-wise addition. Subsequently, a 1 × 1 pointwise convolution is applied to the fused feature to enable inter-channel interaction and dimensional compression:

(8)

To capture long-range global dependencies, PKIBlock introduces a Context Anchor Attention (CAA) branch, which generates channel attention weights through global information aggregation and lightweight mapping. First, the input feature is processed by global average pooling, followed by channel transformation to obtain a compact contextual representation. Subsequently, two depthwise separable convolutions are applied along the horizontal and vertical directions, respectively, to spatially propagate the context vector, enabling effective dissemination of global information.

(9)

Here, and denote the kernel sizes along the horizontal and vertical directions, respectively. This cascaded strip convolution design enables the module to capture spatial contextual information within a large receptive field while maintaining low parameter and computational overhead. Finally, the CAA branch produces a channel attention weight map .

(10)

After obtaining the multi-scale convolution feature and the channel attention weight , PKIBlock integrates them through a gated residual fusion mechanism, resulting in the final density map .

(11)

After being processed by the PKIBlock, the resulting density map serves two purposes. On the one hand, it is fed into a global average pooling layer followed by a fully connected layer to produce the counting output, which is used to determine the number of queries in the subsequent dynamic query selection. On the other hand, the density map is treated as an intermediate feature and forwarded to the CGFE module for further feature enhancement.

Counting-guided feature enhancement module (CGFE)

The CGFE module is an important component of the adaptive query modeling strategy and leverages the spatial prior information embedded in the density map generated by the CCM to enhance feature responses in object-relevant regions of the encoder features. It consists of two components, namely spatial attention and channel attention. Guided by the density map, CGFE employs a dual-attention mechanism to refine feature representations. The overall architecture of CGFE is illustrated in Fig 8:

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Architecture of CGFE.

https://doi.org/10.1371/journal.pone.0355299.g008

Spatial attention: First, the CGFE module employs a spatial attention mechanism to capture important spatial locations in the input feature maps. Let denote the multi-scale feature maps output by the neck, where represents the number of feature levels. The density map is resized to match the spatial resolution of each feature level via a 1 × 1 convolution followed by interpolation, resulting in aligned density features . For each level , the spatial attention weights are computed as follows:

(12)

The spatial attention weights are then applied to the corresponding feature maps via element-wise multiplication, producing spatially enhanced features:

(13)

Channel attention: After spatial attention refinement, CGFE further performs channel attention on the spatially enhanced feature :

(14)

The resulting channel attention weights are then multiplied element-wise with the spatially enhanced features to obtain the final enhanced features at each level:

(15)

Finally, the enhanced features are fed into the Dynamic Query Selection module, where they are used to compute objectness scores for query selection and to generate the content and positional embeddings of the initial queries.

Dynamic query selection

As the final component of the adaptive query modeling strategy, the Dynamic Query Selection module adaptively determines the number of queries and their initial positions for each image based on the predicted object count and the enhanced features. For query content and positional initialization, the initial queries are derived from the enhanced features produced by the CGFE module. Specifically, the multi-scale enhanced features are flattened into a two-dimensional feature sequence , and a feed-forward network is employed to compute objectness scores for each spatial location. The Top-K features with the highest scores are then selected to form a candidate feature set :

(16)

The candidate feature set contains rich semantic information, which is further transformed via a linear projection to generate the content embeddings of the queries:

(17)

Each selected feature point corresponds to a spatial location in the original feature map. Its normalized coordinates , together with predefined anchor box sizes , are used as the initial positional priors to construct reference anchors . These anchors are further refined by predicting offset adjustments through a feed-forward network (FFN), resulting in the final positional priors:

(18)

The resulting dynamic queries are then fed into the decoder for iterative refinement, and the final detection results are produced by the prediction head.

Experiments

Dataset description.

The proposed method is evaluated on three datasets, including CODrone, VisDrone2019, and a self-constructed PV-DV dataset.

CODrone [36] is released by Xiamen University and is designed for oriented object detection in UAV aerial imagery under real-world flight conditions. It contains 10,004 images with annotations covering 12 categories of typical urban and traffic-related objects, including vehicles, traffic facilities, pedestrians, light transportation tools, and vessels. The dataset is split into training, validation, and test sets with a ratio of 5:2:3. Sample images and annotation statistics are illustrated in Fig 9(a).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Datasets label distribution map from the (a) CODrone, (b) VisDrone2019, and (c) PV-DV datasets.

https://doi.org/10.1371/journal.pone.0355299.g009

VisDrone2019 [37], constructed by the AISKYEYE team from Tianjin University, is a large-scale UAV-based benchmark dataset. All images are collected from 14 cities in China and cover diverse scenarios such as traffic intersections, urban environments, and crowded areas. The dataset contains 8,629 images, which are divided into 6,471 training images, 548 validation images, and 1,610 test images. It includes 10 object categories, such as pedestrians, cars, bicycles, tricycles, and buses. In addition to object locations and categories, annotations also include attributes such as occlusion levels and lighting conditions. Sample images and annotation statistics are shown in Fig 9(b).

PV-DV is an infrared photovoltaic defect dataset constructed by our team to support UAV-based inspection and defect detection in large-scale photovoltaic power plants. The dataset initially contains 699 infrared images and is expanded to 1398 images through data augmentation to improve sample diversity and model robustness. It includes four typical defect types: single hot spots, large-area hot spots, abnormal low-temperature regions, and diode short circuits. Specifically, there are 15716 instances of single hot spots, 1336 large-area hot spots, 11804 abnormal low-temperature defects, and 2888 diode short circuits. These annotations comprehensively reflect the diversity and complexity of real-world photovoltaic defect distributions. Based on the width–height distribution of bounding boxes, the dataset is well suited for tiny-object detection. It is divided into training, validation, and test sets with a ratio of 7:2:1, ensuring a balanced distribution of defect categories across subsets. Sample images and annotation statistics are shown in Fig 9(c).

To further clarify the object scale distribution of the CODrone dataset, we analyze the relative area of annotated instances following the COCO and AI-TOD evaluation protocols. As shown in Fig 10, 2.86% of instances fall into the very-tiny category, 9.26% fall into the tiny category, and 54.55% belong to the small-object category.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Object scale distribution of the CODrone dataset.

The categories are defined according to COCO-style area thresholds and AI-TOD-inspired relative-area analysis.

https://doi.org/10.1371/journal.pone.0355299.g010

These results indicate that CODrone is highly dominated by small-scale and tiny-scale objects, which are consistent with real-world UAV aerial imaging scenarios. This distribution further validates the suitability of CODrone for evaluating UAV tiny-object detection methods.

Implementation details

All experiments are conducted on a hardware platform equipped with an Intel Xeon Platinum 8470Q CPU (2.10 GHz) and an NVIDIA RTX 4090 GPU with 24 GB memory. The software environment includes Ubuntu 22.04, Python 3.12, PyTorch 2.3, and CUDA 12.1. During training, a flat-cosine learning rate schedule is adopted, where the first 500 iterations are used for linear warm-up. After the 72nd epoch, strong data augmentation strategies are gradually disabled, and only basic augmentations are retained to facilitate stable convergence. Unless otherwise specified, all hyperparameters follow the default settings of the official implementation. Detailed experimental configurations are provided in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Experimental settings.

https://doi.org/10.1371/journal.pone.0355299.t001

Ablation experiments

To validate the effectiveness of the proposed components, ablation experiments are conducted on the CODrone dataset, and the results are presented in Table 2. Using D-FINE-M as the baseline, we progressively incorporate different components and analyze the resulting changes in detection performance and computational complexity. Unless otherwise specified, all FPS and latency results in the following tables are measured under the same PyTorch FP16 inference setting with batch size 4 and an input resolution of 640 × 640.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Ablation experiments on the CODrone dataset. AQM denotes the adaptive query modeling strategy.

https://doi.org/10.1371/journal.pone.0355299.t002

Specifically, when only the Fixed-Query Self-Attention (FQSA) mechanism is introduced, the AP improves from 16.4 to 17.1, while the number of parameters and computational cost decrease to 18.67M and 55.8 GFLOPs, respectively. Meanwhile, the FPS increases from 118.5 to 133.3, indicating that FQSA improves inference efficiency while still providing stable performance gains. After incorporating the LREA module, the AP further increases to 17.2, with AP_m reaching 17.5. This suggests that expanding the receptive field and enhancing multi-scale feature fusion can effectively improve the representation capability for tiny objects, although the computational cost correspondingly rises to 68.8 GFLOPs and the FPS decreases to 98.7. When only the adaptive query modeling strategy is introduced, the AP reaches 17.0, validating the effectiveness of dynamic query allocation and enhanced feature interaction in improving target matching quality, with an FPS of 104.0. More importantly, combining multiple components further exploits their complementary benefits. When FQSA, LREA, and the adaptive query modeling strategy are jointly integrated, the model achieves the best performance, reaching 17.8 AP, 34.4 AP₅₀, and 17.6 AP_m, while maintaining 85.7 FPS. Compared with the baseline, these results demonstrate consistent improvements in both overall detection accuracy and tiny-object detection performance. Although the multi-scale feature enhancement and dynamic query mechanism introduce additional computational overhead, the model still maintains practical inference speed, reflecting a favorable trade-off between detection accuracy and computational efficiency. It should be noted that, according to the COCO evaluation protocol, AP_s is commonly used to evaluate tiny-object detection performance. However, on the CODrone dataset, the AP_s values are observed to be extremely low and highly unstable, making it difficult to effectively distinguish the performance differences among methods. Therefore, AP_m is adopted in this work as a more stable alternative metric.

To further validate the generalizability of the above findings across different aerial scenarios, additional ablation experiments are conducted on the VisDrone training set to examine the stability and effectiveness of AQF-Net under different data distributions. The experimental results are summarized in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Ablation experiments on the VisDrone dataset. AQM denotes the adaptive query modeling strategy.

https://doi.org/10.1371/journal.pone.0355299.t003

According to Table 3, with the progressive introduction of FQSA, LREA, and the adaptive query modeling strategy, AP steadily increases from the baseline value of 28.7 to 33.1, while AP₅₀ and AP_s reach 54.6 and 21.2, respectively. In terms of inference speed, FQSA increases the FPS from 120.5 to 127.7, further confirming its efficiency advantage. After introducing LREA and the full adaptive query modeling strategy, the FPS decreases to 101.8 and 78.3, respectively, due to the additional feature enhancement and query allocation operations. Nevertheless, the full AQF-Net still maintains real-time inference capability while achieving substantial accuracy gains. These results indicate that each component effectively improves both overall detection performance and tiny-object detection capability, while maintaining a favorable balance between accuracy and efficiency. The ablation results further confirm the effectiveness of the proposed components across different scenarios and demonstrate their complementary roles within the AQF-Net framework.

To evaluate the statistical reliability of the performance improvement, we conduct three independent runs with different random seeds on the VisDrone2019 validation set. The D-FINE baseline and AQF-Net are trained and evaluated under the same experimental settings. We report the mean and standard deviation of AP, AP₅₀, and AP₇₅, and perform a two-sided paired t-test between the two methods.

As shown in Table 4, AQF-Net consistently outperforms the D-FINE baseline across all three runs. The improvements in AP, AP₅₀, and AP₇₅ are 5.27%, 7.20%, and 6.13%, respectively, with all p-values below 0.01. These results indicate that the performance gains of AQF-Net are statistically reliable and not merely caused by random seed variation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Statistical significance analysis on the VisDrone2019 validation set over three independent runs.

https://doi.org/10.1371/journal.pone.0355299.t004

Comparative experiments

The comparative results on the VisDrone validation set are reported in Table 5. Overall, conventional general-purpose detectors, such as YOLO and DETR, suffer from evident performance bottlenecks in this challenging aerial scenario. Even when the model scale is enlarged or deformable attention mechanisms are introduced, the improvement in detection accuracy remains limited and is often accompanied by a substantial increase in model parameters and computational cost. In contrast, methods specifically designed for UAV tiny-object detection demonstrate more competitive performance. Approaches such as QueryDet, SINextNet, SOD-YOLO, UAV-DETR, and SO-DETR improve detection accuracy by introducing mechanisms including dense sampling, hierarchical feature reweighting, and query-level modeling. However, some of these methods, such as UAV-DETR, SO-DETR, and FBRT-YOLO-X, rely on highly complex Transformer decoding structures or multi-branch feature interaction designs, resulting in a significant increase in parameter size and computational complexity. This imposes considerable deployment pressure in practical UAV platforms and real-time application scenarios. Overall, existing methods still struggle to achieve an ideal balance between detection accuracy and computational efficiency. This further highlights the necessity and research value of structured modeling and query optimization for aerial tiny-object detection under controllable computational overhead. The proposed AQF-Net-M achieves the best performance on the VisDrone validation set, reaching 33.4 AP, 55.0 AP₅₀, and 33.9 AP₇₅, while maintaining a controllable model size of 22.70M parameters and a computational cost of 76.7 GFLOPs. These results significantly outperform conventional general-purpose detectors as well as most UAV-specific methods, demonstrating the effectiveness and superiority of the proposed adaptive query modeling and feature interaction mechanisms in complex aerial tiny-object detection scenarios.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Comparison of methods on the VisDrone validation set.

https://doi.org/10.1371/journal.pone.0355299.t005

To evaluate the performance of different detection methods on photovoltaic array defect detection, comparative experiments are conducted on the PV-DV dataset, and the results are presented in Table 6. The results show that AQF-Net-M achieves the best performance in terms of AP, AP₅₀, and AP₇₅, reaching 58.7%, 85.5%, and 62.3%, respectively, significantly outperforming the other competing methods. In terms of inference efficiency, AQF-Net-M achieves 70.2 FPS with a latency of 14.24 ms, indicating that the improved accuracy is obtained at the cost of additional computational overhead. As a lightweight version of AQF-Net-M, AQF-Net-S is designed for resource-constrained scenarios by reducing the number of internal channels in the model. Despite its lower complexity, it still demonstrates strong detection performance, achieving 57.5% AP, 84.5% AP₅₀, and 61.6% AP₇₅, while improving the inference speed to 109.3 FPS with a latency of 9.15 ms. Overall, the AQF-Net series outperforms other mainstream detection methods on the PV-DV dataset, further validating the effectiveness and application potential of the proposed method for photovoltaic array defect detection.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Comparison of methods on the PV-DV dataset.

https://doi.org/10.1371/journal.pone.0355299.t006

Roles of key innovations

Effect of adaptive query allocation.

To investigate the impact of different configurations of the dynamic query mechanism within the adaptive query modeling strategy, systematic ablation experiments are conducted on the CODrone dataset based on the D-FINE-M architecture. While keeping the core idea of dynamic query allocation and the overall decoding pipeline unchanged, these experiments only modify the Categorical Counting Module through lightweight redesign or structural enhancement. The experimental results are summarized in Table 7. For clarity, the final configuration with the lightweight counting module and PKIBlock is referred to as the proposed adaptive query modeling strategy in AQF-Net, whereas the other two configurations are intermediate variants used only for ablation analysis.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Ablation experiments on the dynamic query mechanism.

https://doi.org/10.1371/journal.pone.0355299.t007

As shown in Table 7, after introducing the dynamic query mechanism into the D-FINE-M baseline, the overall detection performance improves steadily, but the number of parameters and computational cost also increase significantly. Based on this observation, we further perform structural ablation analysis on the Categorical Counting Module. The results show that the lightweight design effectively reduces computational overhead while maintaining detection accuracy, indicating that the original counting module in the dynamic query mechanism contains a certain degree of structural redundancy. Furthermore, after incorporating PKIBlock, the model achieves the best performance in terms of AP, AP₅₀, and AP₇₅, while introducing no significant additional computational cost compared with the baseline. These results demonstrate that proper optimization of the internal structure of the dynamic query mechanism helps achieve a better balance between detection accuracy and computational efficiency.

To verify whether the improvement of AQM comes from density-aware query allocation rather than merely varying the number of decoder queries, we conducted controlled ablation experiments on the VisDrone2019 validation set. In particular, we introduce a random query masking baseline. In this setting, the number of decoder queries is randomly sampled from the same candidate query list, i.e., [300, 330, 360, 400], while the CCM modules are retained. Therefore, random query masking preserves the effect of using dynamic query numbers but removes the density-aware allocation rule.

As shown in Table 8, compared with the D-FINE baseline using a fixed number of 300 queries, introducing CCM while keeping the query number unchanged improves AP from 28.7% to 31.5%, AP₅₀ from 48.0% to 52.1%, and AP₇₅ from 28.4% to 31.8%. This indicates that the counting-guided feature enhancement provides effective density-aware priors for improving feature representation. However, when the number of queries is randomly selected from [300, 330, 360, 400], the model achieves only 30.2% AP, which is lower than the fixed-query CCM variant and clearly inferior to the full AQM. This suggests that varying the number of decoder queries alone does not consistently improve detection performance and may even introduce mismatches between scene density and query allocation. In contrast, the full AQM achieves the best performance, with 33.1% AP, 54.6% AP₅₀, and 33.7% AP₇₅. These results demonstrate that the improvement of AQM is not merely caused by changing or reducing the number of queries, but mainly benefits from the density-aware query allocation guided by the CCM, which adaptively aligns the query budget with scene-level object density and alleviates redundant background predictions and query competition in UAV tiny-object detection.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Ablation study of adaptive query allocation on the VisDrone2019 validation set.

https://doi.org/10.1371/journal.pone.0355299.t008

To further evaluate the reliability of the Categorical Counting Module (CCM), we analyze its density prediction and query allocation behavior on the VisDrone2019 validation set. According to the number of annotated instances in each image, the ground-truth density levels are divided into four categories: sparse (≤10 objects), moderate (11–60 objects), dense (61–150 objects), and very dense (>150 objects). These four density levels are mapped to 300, 330, 360, and 400 decoder queries, respectively.

As shown in Table 9, CCM achieves 81.75% density classification accuracy and 100.00% adjacent accuracy. The under-allocation and over-allocation rates are 10.04% and 8.21%, respectively, while the query MAE and RMSE are 5.91 and 13.96. The confusion matrix in Fig 11 further shows that all misclassified samples fall into adjacent density categories, and no severe cross-level prediction error occurs. These results indicate that CCM provides stable density-aware guidance for dynamic query allocation, even when the predicted density level is not exactly correct.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Reliability analysis of CCM on the VisDrone2019 validation set.

https://doi.org/10.1371/journal.pone.0355299.t009

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Normalized confusion matrix of CCM density prediction on the VisDrone2019 validation set.

https://doi.org/10.1371/journal.pone.0355299.g011

Effect of query-space resolution in FQSA

To investigate whether the fixed low-resolution query space in FQSA affects the localization accuracy of tiny objects, we conduct a sensitivity analysis by varying the query resolution while keeping all other settings unchanged. Specifically, we evaluate three query resolutions (8, 16, and 24) and compare them with the original D-FINE baseline.

As shown in Table 10, introducing FQSA with an excessively small query resolution (8) slightly decreases detection performance, particularly under the stricter AP₇₅ metric, indicating that excessive query compression may weaken fine-grained localization. Increasing the query resolution to 16 significantly improves AP, AP₅₀, and AP₇₅, achieving the best overall performance with 17.2% AP and 16.0% AP₇₅. Although a larger query resolution (24) further improves AP₇₅ to 16.3%, the overall AP drops slightly while the computational cost increases from 55.8 GFLOPs to 57.1 GFLOPs. These results demonstrate that FQSA effectively balances efficiency and localization accuracy by performing attention computation in a fixed low-resolution query space. The default setting of query resolution 16 provides the most favorable trade-off between detection performance and computational complexity, and is therefore adopted in the final model.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Effect of query resolution in FQSA on the VisDrone2019 validation set.

https://doi.org/10.1371/journal.pone.0355299.t010

Effect of LREA

To further verify the effectiveness and distinctiveness of the proposed LREA module at the neck feature enhancement stage, we conduct comparative experiments on the CODrone dataset. The LREA module at the position shown in Fig 1 is replaced with several representative feature enhancement and large-kernel context modules, including RGCSPELAN, CSSC, CFBlock, FCM, FMB, InceptionNeXt Block, CNCM, LKA, LSKA, and StripConv. In addition, SDBConv is evaluated separately to examine the contribution of the proposed sparse decomposed broad convolution, while the full LREA is used to evaluate the effect of combining SDBConv with bidirectional selective aggregation. All modules are inserted at the same position and trained under identical settings. The experimental results are summarized in Table 11.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Comparison with other feature enhancement modules.

https://doi.org/10.1371/journal.pone.0355299.t011

Overall, different neck enhancement modules exhibit substantial differences in the trade-off between detection accuracy and computational efficiency. Lightweight modules such as RGCSPELAN, FCM, FMB, and InceptionNeXt Block reduce the number of parameters or GFLOPs, but their performance improvements are limited. For example, RGCSPELAN reduces the computational cost to 33.9 GFLOPs but decreases AP from 16.4% to 15.5%, while InceptionNeXt Block also results in lower AP and AP₅₀ than the baseline. This suggests that excessive lightweight design may weaken fine-grained feature representation for UAV tiny-object detection. Existing large-kernel or directional context modules also show different behaviors. StripConv and LSKA reduce computational complexity to 45.5 and 25.6 GFLOPs, respectively, but their AP drops to 16.1% and 15.1%, indicating that simply introducing strip convolution or separable large-kernel attention does not necessarily improve tiny-object perception. LKA achieves better efficiency with 26.4 GFLOPs and improves AP to 16.7%, but its AP₇₅ remains lower than the baseline, suggesting that fixed large-kernel aggregation may be insufficient for precise localization. In contrast, SDBConv improves AP to 16.8% and AP₅₀ to 32.2%, demonstrating that decomposed broad receptive-field modeling can provide more effective contextual representation than conventional strip or large-kernel designs. It is worth noting that CNCM achieves relatively high performance, with 17.0% AP, 32.5% AP₅₀, and 16.0% AP₇₅. However, this improvement comes at a considerable computational cost, as its parameter count and complexity increase sharply to 49.32M and 196.8 GFLOPs, respectively. By comparison, the proposed LREA achieves the best overall performance on the CODrone dataset, reaching 17.2% AP, 32.7% AP₅₀, and 16.4% AP₇₅, while keeping the parameter count and computational cost at 21.55M and 68.8 GFLOPs. Compared with SDBConv alone, LREA further improves AP, AP₅₀, and AP₇₅, indicating that the BSA mechanism provides additional benefits by adaptively selecting complementary contextual responses across different branches. These results demonstrate that LREA achieves a more favorable balance between accuracy and efficiency, and its advantage comes not merely from enlarging the receptive field, but from combining decomposed broad receptive-field modeling with adaptive multi-scale and directional feature aggregation.

To provide a more intuitive comparison of the impact of different neck modules on feature representation, Fig 12 presents the feature maps produced by several representative feature enhancement modules on the CODrone dataset, including RGCSPELAN, CSSC, FMB, InceptionNeXt Block, LKA, StripConv, SDBConv, and the proposed LREA. As shown in the figure, the model exhibits noticeable differences in target perception as the neck enhancement module changes. Compared with conventional lightweight enhancement modules, large-kernel or strip-convolution-based modules can capture broader contextual information, but their feature responses are still relatively scattered or affected by background textures. SDBConv strengthens the responses around object regions by introducing decomposed broad receptive-field modeling. Furthermore, when the full LREA is adopted, the resulting feature maps show clearer target contours and richer local details. This advantage is especially evident for tiny objects, where LREA not only expands the receptive field but also adaptively integrates multi-scale and directional contextual information through BSA, thereby enhancing the discriminability of target features.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Feature maps processed by different feature enhancement modules.

https://doi.org/10.1371/journal.pone.0355299.g012

Training stability analysis

To further analyze whether the adaptive query modeling strategy affects training stability, we visualize the training loss curves and the average query number evolution, as shown in Fig 13.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 13. Training stability analysis of AQF-Net.

(a) Training loss curves of D-FINE and AQF-Net. (b) Average query number evolution of AQF-Net during training.

https://doi.org/10.1371/journal.pone.0355299.g013

Fig 13(a) shows that both D-FINE and AQF-Net exhibit smooth and stable convergence under the same training schedule. Compared with the D-FINE baseline, AQF-Net achieves a consistently lower training loss and converges to a lower final loss value, indicating that the proposed modules do not introduce optimization instability. Fig 13(b) further illustrates the evolution of the average number of decoder queries during training. The average query number increases gradually in the early training stage and then becomes stable in the later stage, with only small fluctuations. This suggests that the Categorical Counting Module provides stable density-aware guidance for query allocation rather than causing abrupt query oscillations. Overall, these results demonstrate that AQF-Net maintains stable convergence behavior while dynamically adjusting the query budget according to scene density.

Visualization analysis

D-FINE-M AQF-Net-M.

Fig 14 presents a comparison of the feature response heatmaps of D-FINE-M and AQF-Net-M on representative samples from the VisDrone and CODrone datasets. It can be observed that, under complex background conditions, the activation regions of D-FINE-M are relatively scattered, and its responses to some tiny objects are not sufficiently prominent, making the model more susceptible to interference from background textures and noise. In contrast, AQF-Net-M produces more concentrated and clearer high-response distributions in object regions. Especially in scenarios with densely distributed tiny-scale objects, the model is able not only to generate high-confidence activations at target locations, but also to effectively suppress responses in irrelevant background regions. These results indicate that, after incorporating LREA and FQSA, the model improves its multi-scale contextual modeling capability while maintaining computational efficiency, thereby enabling it to focus more consistently on tiny-object regions. In addition, the adaptive query modeling strategy, through counting-guided feature enhancement and query selection, allows the model to adaptively adjust its attention regions in scenarios with large variations in object density, resulting in more discriminative response distributions in the heatmaps.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 14. Heatmap comparison between D-FINE-M and AQF-Net-M.

https://doi.org/10.1371/journal.pone.0355299.g014

Detection comparison

D-FINE-M AQF-Net-M.

Fig 15 presents a visual comparison of the detection results of D-FINE-M and AQF-Net-M on the VisDrone and PV-DV datasets. As can be observed, in the VisDrone scenario, AQF-Net-M is able to detect more densely distributed tiny objects while significantly reducing both missed detections and false positives. On the PV-DV dataset for photovoltaic defect detection, AQF-Net-M not only maintains a high recall rate but also achieves more accurate defect localization results. This advantage can be mainly attributed to the ability of AQF-Net-M to adaptively adjust the number of queries according to object density, thereby improving detection stability and robustness in different scenarios.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 15. Detection results comparison on VisDrone and PV-DV datasets.

https://doi.org/10.1371/journal.pone.0355299.g015