DSENet addresses insufficient feature extraction for small objects. Drawing from RepViT [1]’s reparameterization structure, we propose a progressive dense SE deployment strategy specifically for small object detection, distinct from RepViT’s uniform sparse design (∼50% coverage) for general tasks: we systematically escalate SE module density from 50% in shallow layers to 90% in deep layers, implementing dense channel recalibration where small objects are compressed to 1–3 pixels in deep feature maps (1/32 resolution). This density progression, driven by small object feature sparsity analysis rather than NAS-based general metric optimization, specifically addresses constraints of extreme scale imbalance in small object detection. As shown in Fig. 2, DSENet constructs Dense Attention Units (DAU) as basic blocks, progressively extracting and optimizing features through four stages.

Visualization Verification: Fig. 3 compares feature activation patterns of Baseline, RepViT, and DSENet across four depth levels (P2-P5). DSENet demonstrates superior small-object activation in deep layers (Stage 4, P5/32), achieving +2585.9% SNR improvement over RepViT (0.0228→0.6127), validating progressive dense SE deployment for small object feature sparsity.

The core of DSENet lies in the design of DAU units. For input feature map X∈RC×H×W, DAU processes through spatial mixing and channel mixing steps. Spatial mixing adopts a multi-branch structure to enhance representation, combined with SE [13] module weighting: (1)X′=SE(BN(Conv3×3DW(X)+Conv1×1DW(X)+X))

Channel mixing achieves cross-channel information interaction through residual structure: (2)Y=X′+Conv1×1down(GELU(Conv1×1up(X′)))

This design combines efficient spatial feature extraction with dense channel attention, learning robust feature representations during training and can be reparameterized into a single convolution layer during inference, significantly enhancing the perception capability of key features for small objects.

To overcome the information loss problem in traditional feature fusion strategies, we design the Adaptive Gated Fusion module (AGF). Unlike fixed-weight attentional fusion, we propose a learnable gating parameter mechanism: the λ parameter in Eq. (5) is an nn.Parameter tensor optimized end-to-end, not a hyperparameter. It converges to scale-specific values (0.42±0.08→0.71±0.15 from shallow to deep), automatically adapting to background dominance (>99.2% in P5). Unlike NAS methods requiring ∼500 GPU-hours for exhaustive search, AGF achieves comparable performance via gradient descent in <3 GPU-hours (200× efficiency). Through end-to-end trained λ (Eq. (5)), we dynamically modulate multi-scale feature contributions, learning optimal fusion strategies for UAV small object detection. As shown in Fig. 4, AGF adaptively fuses features through precise gating.

AGF consists of three components: dual-branch channel gating unit, enhanced local gating unit, and adaptive fusion mechanism. The dual-branch channel gating unit employs global pooling in parallel to capture channel statistics, processed through adaptive one-dimensional convolution: (3)k=|log2⁡(C)+bγ|oddwhere k is the convolution kernel size, adaptively calculated based on the number of channels C, to efficiently capture cross-channel interactions.

The enhanced local gating unit integrates H-W decomposition attention and spatial attention, combined through learnable parameters (γ, δ, ε are nn.Parameter tensors optimized end-to-end, enabling task-specific weighting of heterogeneous attention pathways): (4)Y=γ⋅(X⊙Ah⊙Aw)+δ⋅(X⊙Aspatial)+ε⋅X

The core gating fusion mechanism enhances input features through dual-branch attention and introduces learnable parameter λ as a “gate” for adaptive weighting: (5)Xfused=X1enhanced+λ⋅X2

Through this design, AGF achieves comprehensive enhancement of features, addressing the information dilution problem and providing information-rich fused features.

AGF vs. NAS: While employing classic operations, AGF embeds task-specific inductive bias (scale imbalance) into differentiable parameters (λ, γ, δ, ε) rather than exhaustive search. This achieves: (1) 200× efficiency (3 vs. 500 GPU-hours); (2) interpretability (learned λ gradient 0.42→0.71 reveals scale adaptation); (3) cross-dataset robustness (VisDrone/CODrone consistent gains vs. NAS overfitting risk).

Scale-adaptive mechanism:λ parameters converge to distinct value ranges across pyramid levels: P2 (0.42±0.08) preserves shallow details, P3-P4 (0.58±0.12) balances multi-scale information, P5 (0.71±0.15) suppresses background (>99.2%). This progression (0.42→0.71) correlates with feature sparsity, validating automatic adaptation to UAV scale imbalance.

To enhance weak feature representation of small objects, we design the Channel-Spatial Attention module (CSA). Unlike simple weighted combinations with fixed coefficients, CSA employs nn.Parameter tensors (α, β in Eq. (8)) optimized end-to-end. They converge to task-specific values (α=0.62±0.08, β=0.38±0.05) via gradient descent, automatically discovering channel-spatial balance. The parallel architecture preserves complete feature distributions (vs. cascaded information loss), exhibiting depth-dependent patterns (shallow: 0.58/0.42; deep: 0.65/0.35) that validate scale-aware adaptation. Addressing limitations of single-dimension designs (e.g., AFGCAttention [14]) and sequential cascading (e.g., CBAM), we propose information flow parallelization: channel and spatial attention synchronously model in original feature space, eliminating cascaded bottlenecks. Learnable fusion coefficients (Eq. (8)) dynamically balance dual-dimension contributions, achieving 1.24× variance improvement and spatial focusing for small objects (<0.1% spatial occupancy). As shown in Fig. 5, CSA enhances features through dual-pathway parallelism.

This module receives feature map X∈RC×H×W, processing it through parallel channel attention branch and spatial attention branch. The channel attention branch employs global pooling to capture statistics, capturing channel interactions through dynamic one-dimensional convolution, with kernel size adaptively calculated based on the number of channels: (6)k=|log2⁡(C)+bγ|odd

The spatial attention branch identifies important spatial regions, generating attention maps through channel dimension pooling and convolution: (7)Aspatial=σ(Conv7×7(cat([AvgPool(X),MaxPool(X)])))

The core innovation of CSA lies in its adaptive fusion mechanism, dynamically weighting the two types of attention through learnable parameters: (8)Xenhanced=(X⊙Achannel)⋅α+(X⊙Aspatial)⋅β

This design enables the module to precisely recalibrate features based on comprehensive information, significantly improving the perception and discrimination capabilities of weak features.

Fig. 6 validates CSA’s feature enhancement on dense small object scenes. Baseline model shows scattered channel attention (σ = 0.149) and background-focused spatial attention. CSA enhancement increases channel variance to 0.184 (1.24× improvement) while spatial attention precisely focuses on small object regions, demonstrating synergistic dual-pathway optimization.

We conducted comprehensive comparative experiments under identical settings to verify each component’s effectiveness. Table 3 summarizes the detailed comparative experimental results.

Specifically, for baseline comparison, we selected YOLOv5 [17], YOLOv6 [18], YOLOv8 [7], and YOLOv10n [3], as well as Transformer-based RT-DETR [19] to ensure comprehensive evaluation across architectural paradigms. For backbone networks, we compared HGNetV2 [19], EfficientViT [20], ConvNeXtV2, and FasterNet [21] against our DSENet. For feature fusion modules, we evaluated CARAFE [22] and CGAFusion alongside our AGF. For attention mechanisms, we compared MSGA [23], SimAM [24], and CPCA with our CSA. Additionally, to comprehensively address the Reviewer’s concern, we evaluated RT-DETR integrated with AIFI-SHSA [25] and Cascaded Group Attention [20] to compare Transformer-based architectures with traditional attention mechanisms.

The baseline comparison validates YOLOv10n’s suitability as the foundational architecture (mAP50: 32.86%). DSGF-Net achieved mAP50 of 37.98% (+5.12 pp) and mAP50-95 of 22.27% (+3.54 pp), demonstrating substantial improvements. We further evaluated RT-DETR with AIFI-SHSA (29.96% mAP50) and CGA (29.62% mAP50), both marginally improving vanilla RT-DETR (29.78%, <0.2 pp gain) yet 7.85–8.35 pp lower than DSGF-Net (37.97%). This validates: (1) generic attention additions provide minimal gains for Transformer-based small object detection, confirming O(n2) self-attention limitations on high-resolution UAV imagery; (2) DSGF-Net’s task-specific design achieves 27.5% relative improvement using only 33.8% parameters (6.67 vs. 19.71 M), demonstrating specialized architectural superiority. RT-DETR, despite 19.88M parameters (8.8× YOLOv10n), achieves only 29.78% mAP50 (8.19 pp lower than DSGF-Net), validating DSGF-Net’s parameter efficiency: 33.5% parameters achieve 27.5% relative accuracy improvement.

Component-wise analysis reveals distinct contributions: DSENet as the backbone yields the most significant gain (mAP50: 36.98%, +4.12 pp), attributed to its structure reparameterization and optimized attention distribution, outperforming alternatives like FasterNet and ConvNeXtV2. The AGF module achieves mAP50 of 34.04% (+1.18 pp) through its adaptive gating strategy, surpassing CARAFE and CGAFusion while maintaining computational efficiency. The CSA module attains mAP50 of 33.65% and F1 score of 38.32%, enhancing feature representation without additional computational overhead.

Table 4 presents the multi-dataset comparison results. To comprehensively evaluate the generalization ability and robustness of the DSGF-Net model, we deployed it on two UAV aerial datasets with significantly different characteristics—VisDrone and CODrone—and compared its performance with the baseline model YOLOv10n.

On the classic VisDrone dataset, DSGF-Net demonstrated excellent performance advantages, achieving an mAP@0.5 of 37.97%, a significant improvement of 5.11 percentage points over the baseline model YOLOv10n. On the more challenging CODrone dataset, DSGF-Net still achieved an mAP@0.5 of 24.28%, an improvement of 2.36 percentage points over the baseline. This ability to maintain stable performance gains on datasets of different complexity and characteristics fully demonstrates DSGF-Net’s strong robustness and generalization potential, proving its universal improvement capability to address different real-world challenges.

Robustness Evaluation: To validate DSGF-Net’s adaptability in real UAV scenarios, we evaluated performance under different flight altitudes and motion blur conditions. As shown in Fig. 7, by adjusting validation image sizes to simulate altitude variations (50 m/100 m/200 m corresponding to 960/640/320 pixels), DSGF-Net maintains significant advantages over baseline at all altitudes (42.19%/37.97%/21.09% vs. 37.51%/32.86%/17.74%). In motion blur tests (blur kernel 0/5/11), DSGF-Net similarly demonstrates stronger anti-interference capability (37.97%/35.12%/23.37% vs. 32.86%/30.54%/20.54%). Notably, under extreme conditions (200 m altitude or kernel = 11 blur), DSGF-Net still maintains 3–4 percentage point performance advantages, attributed to DSENet’s dense attention mechanism and AGF’s adaptive fusion enhancing feature robustness, validating the practical value of the proposed architecture in complex environments.

To explore the synergistic gains between modules and validate the reasonability of the overall architecture, we conducted ablation experiments. Table 5 presents the comprehensive ablation study results with computational complexity analysis.

As illustrated in Fig. 8, the results indicate that the synergy between modules is not a simple performance addition, but follows a clear functional complementary logic.

DSENet (B) serves as the performance cornerstone; its presence or absence is key to determining the model’s performance level. All combinations containing DSENet significantly outperform other combinations, again proving that high-quality initial feature extraction is a prerequisite for achieving excellent detection accuracy.

AGF (C) module plays a crucial “bridge” role in synergy. With DSENet providing high-quality features, the introduction of AGF (B+C combination) can further bring significant performance improvements, increasing mAP50 from 36.98% to 37.61%. This strongly proves that AGF’s adaptive gating mechanism can efficiently integrate multi-scale feature flows produced by DSENet, achieving a “1+1>2” amplification effect. Notably, the A+C combination (34.15%) performs slightly worse than using C alone (34.04%), exhibiting negative synergy. In-depth analysis reveals this stems from hierarchical dependency of feature quality: AGF’s adaptive gating mechanism is designed for fine-grained weight allocation among high-quality multi-scale features; when input features are insufficient, gating parameter λ cannot effectively function, introducing parameter redundancy instead. Meanwhile, CSA’s attention recalibration on weak features easily falls into local optima, causing interference between optimization objectives of both modules. This finding emphasizes the systematic principle of architecture design: feature enhancement modules must build upon high-quality feature extraction foundations to achieve synergistic gains.

Finally, the complete DSGF-Net (A+B+C) model reached the peak of performance, benefiting from the functional complementarity and progressive optimization of the three modules as an organic whole. In this architecture, DSENet is responsible for building a high-quality feature foundation; AGF serves as the central hub, optimizing and integrating multi-scale information flows; finally, the lightweight CSA performs final fine-tuning enhancement on the already highly optimized feature maps. This design with clear division of labor and high synergy makes the comprehensive performance of the entire system exceed the effect of simple addition of various parts, fully validating the scientific nature and advancement of our overall architectural design.

Computational complexity and performance trade-off analysis: From Table 5’s complexity metrics, complete DSGF-Net increases 193.8% parameters (2.27M→6.67M) and 174% computation (6.5→17.8 GFLOPs) compared to baseline, yet achieves 5.11 percentage points mAP50 improvement (15.6% relative gain), with efficiency ratio reaching 2.6. In-depth analysis reveals rationality of complexity increase: (1) DSENet is core contributor: though adding 188.1% parameters alone, it brings 4.12 pp accuracy gain with efficiency ratio 2.2, validating cost-effectiveness of progressive dense SE deployment for small object feature sparsity; (2) Lightweight modules’ efficient synergy: CSA (+2.6% params, ratio 30.4) and AGF (+1.8% params, ratio 65.6) achieve significant gains with minimal parameter overhead, proving importance of refined design; (3) Real-time capability guarantee: despite FPS dropping from 149.9 to 95.0 (36.6% decrease), it far exceeds real-time detection requirements (>>30 FPS). Compared to RT-DETR (19.88M params, 96.0 FPS, 29.78% mAP50), DSGF-Net achieves 8.19 pp higher accuracy with 1/3 parameters and comparable inference speed, fully demonstrating superior balance among accuracy-efficiency-parameters of proposed architecture.