Hierarchical Extended Path Aggregation Network (HEPAN)

• HEPAN is a specialized neck module for HierLight-YOLO that fuses multi-scale features to enhance detection of extremely small UAV targets.
• It employs hierarchical cross-level connections and 1×1 channel compression layers to preserve shallow spatial detail while integrating deep semantic context.
• Empirical evaluations on VisDrone2019 show HEPAN improves AP by 0.8 over PANet with modest increases in parameters and FLOPs, demonstrating its effectiveness for UAV detection.

Hierarchical Extended Path Aggregation Network (HEPAN) is the neck module of HierLight-YOLO, proposed for multi-scale feature fusion in UAV imagery where detection is dominated by extremely small, low-contrast targets embedded in cluttered backgrounds. Within the standard Backbone $\rightarrow$ Neck $\rightarrow$ Head decomposition, HEPAN is positioned between backbone feature extraction and a four-scale detection head, and is presented as an extension of conventional FPN/PAN-style necks. Its defining characteristics are hierarchical cross-level connections, $1\times1$ channel-compression layers at each pyramid level, and dense or residual-style skip links intended to strengthen feature transmission and gradient flow while preserving shallow spatial detail alongside deep semantic context (Chen et al., 26 Sep 2025).

1. Problem setting and design rationale

HEPAN is motivated by the failure mode of real-time detectors in UAV photography: objects are frequently smaller than $32\times32$ pixels, sometimes as small as $4\times4$ pixels, and appear amid complex backgrounds, severe scale variation, and low object contrast. The paper explicitly argues that YOLO-style detectors perform well on larger objects but exhibit higher false negative rates when small drone-view targets dominate the scene. In that framing, the limitation is not only backbone capacity but also the neck’s inability to fuse shallow detail and deep semantics adequately (Chen et al., 26 Sep 2025).

The proposed response is a neck architecture that preserves high-resolution spatial detail from shallow layers while still injecting coarse semantic context from deeper layers. The paper characterizes this as an extension of conventional PANet through two central ideas: intermediate convolutional blocks with residual connections to strengthen gradient propagation, and bidirectional multi-scale feature fusion that preserves both high-resolution details and rich semantic information. More specifically, HEPAN is intended to address insufficient preservation of shallow high-resolution features, limited cross-level interaction beyond adjacent scales, weaker gradient propagation through deeper fusion paths, and a suboptimal balance between semantic abstraction and spatial precision for tiny aerial targets (Chen et al., 26 Sep 2025).

This emphasis places HEPAN within a specific subclass of neck designs: it is not described as a general-purpose replacement for all pyramidal fusion systems, but as a topology specialized for UAV small-object detection, where the preservation of fine-grained cues is treated as a first-order architectural constraint.

2. Position within HierLight-YOLO and feature hierarchy

HierLight-YOLO follows a standard decomposition in which the backbone performs feature extraction, the neck is HEPAN, and the head contains four detection layers of different scales. The backbone includes Conv, C2f, and lightweight IRDCB modules; the neck is HEPAN; and the final detector adds a dedicated high-resolution branch for tiny objects. The paper states explicitly that “The Neck uses the HEPAN to fuse multi-scale features, enhancing small object detection. The Head contains four detection layers of different scales…” (Chen et al., 26 Sep 2025).

HEPAN operates on four backbone outputs, $P2$ through $P5$. At the training resolution of $640\times640$, the paper explicitly identifies $P2$ as $160\times160$ and $\rightarrow$0 as $\rightarrow$1, implying the standard YOLO pyramid:

• $\rightarrow$2 with stride $\rightarrow$3
• $\rightarrow$4 with stride $\rightarrow$5
• $\rightarrow$6 with stride $\rightarrow$7
• $\rightarrow$8 with stride $\rightarrow$9

This four-level hierarchy is architecturally important because the paper contrasts it with standard YOLOv8 necks that typically operate over $1\times1$0–$1\times1$1, with $1\times1$2 appearing only in $1\times1$3 variants. HEPAN is therefore built around a broader pyramid from the outset, not by merely appending a high-resolution output at the end (Chen et al., 26 Sep 2025).

The extra tiny-object branch is attached after HEPAN rather than being generated wholly inside it. Specifically, the branch takes the fused $1\times1$4 output from the neck, upsamples it to $1\times1$5, and concatenates it with shallow backbone $1\times1$6 features. This creates a hybrid neck–head connection in which HEPAN supplies semantically enriched $1\times1$7 features while the head explicitly reintroduces high-resolution shallow detail (Chen et al., 26 Sep 2025).

3. Internal structure and feature-fusion mechanism

A defining component of HEPAN is Hierarchical Feature Channel Compression (HFCC). Each incoming backbone feature from $1\times1$8 to $1\times1$9 is first projected by a $32\times32$0 convolution:

$32\times32$1

The paper motivates this operation using the Information Bottleneck Principle, interpreting the projection as a way to compress redundant information while preserving task-relevant semantics. The only exact channel example given is that $32\times32$2 is compressed from $32\times32$3 channels to $32\times32$4 channels. The associated mutual-information expressions,

$32\times32$5

are presented as conceptual justification rather than as an explicit training objective (Chen et al., 26 Sep 2025).

After HFCC, HEPAN performs bidirectional and hierarchical fusion. The paper describes the neck as extending PANet’s bidirectional flow philosophy while adding dense cross-layer skip connections. Its intended behavior comprises three coupled processes: a top-down path that propagates deep semantic features toward finer scales, a bottom-up path that sends localization- and detail-enhanced features back toward coarser scales, and cross-layer dense skip fusion that introduces more nonlocal scale-to-scale interactions than standard adjacent-level routing. The architecture is therefore not limited to the sparse top-down/bottom-up pattern of ordinary PAN/FPN-style necks (Chen et al., 26 Sep 2025).

Dense skip connections are justified through an explicit gradient-flow expression:

$32\times32$6

The intended interpretation is that multiple downstream fusion nodes contribute gradients back to a shallow feature through dense cross-layer paths. The paper uses this to argue that HEPAN preserves high-resolution details in deep feature computation, mitigates gradient vanishing, and particularly benefits small-object parameter updates (Chen et al., 26 Sep 2025).

The text also states that HEPAN introduces additional $32\times32$7 convolutional layers in the neck network and intermediate convolutional blocks with residual connections. However, it does not provide a node-by-node algebraic schedule for every fusion stage, nor does it explicitly identify IRDCB as the internal fusion block of HEPAN. The safest reading is that HEPAN defines the fusion topology and includes extra $32\times32$8 and residual fusion blocks, while later lightweight substitutions such as IRDCB belong to the broader implementation of HierLight-YOLO rather than to the defining abstraction of HEPAN itself (Chen et al., 26 Sep 2025).

4. Mathematical interface to the detection heads

The most explicit fusion equation in the paper concerns the tiny-object branch that extends the conventional $32\times32$9–$4\times4$0 detector to a four-scale $4\times4$1–$4\times4$2 hierarchy. The paper defines nearest-neighbor upsampling $4\times4$3 with scale factor $4\times4$4, mapping

$4\times4$5

In practice, $4\times4$6 is applied to the neck’s $4\times4$7 output, transforming an $4\times4$8 feature map into a $4\times4$9 feature map (Chen et al., 26 Sep 2025).

The resulting upsampled feature is then concatenated with backbone $P2$0:

$P2$1

This equation is important because it makes explicit how HEPAN interfaces with the dedicated tiny-object path. The branch is not simply another internal HEPAN node; it is an extension built on top of HEPAN’s fused $P2$2 representation and backbone $P2$3 detail. The final detector therefore contains four detection layers:

• $P2$4
• $P2$5
• $P2$6
• $P2$7

The paper does not define any HEPAN-specific loss term. Training uses the standard YOLOv8 anchor-free detection formulation and is evaluated with standard COCO metrics, including $P2$8, $P2$9, and $P5$0 for small objects smaller than $P5$1 (Chen et al., 26 Sep 2025).

A plausible implication is that HEPAN is intended to improve performance entirely through representational and routing changes in the neck, rather than through auxiliary supervision or explicit cross-scale consistency losses.

5. Empirical performance, implementation context, and trade-offs

The most direct experimental evidence for HEPAN is the neck comparison on VisDrone2019-val, where it is evaluated against PANet and BiFPN within the same framework. The paper reports the following values (Chen et al., 26 Sep 2025):

Neck structure	$P5$2	Params / FLOPs
PANet	44.1	10.6M / 36.7G
BiFPN	44.3	10.7M / 37.1G
HEPAN	44.9	11.3M / 38.1G

These results support the paper’s central empirical claim: HEPAN improves $P5$3 by $P5$4 over PANet and by $P5$5 over BiFPN, at the cost of a moderate increase in parameters and FLOPs. The same section states that HEPAN alone improves detection over the conventional feature pyramid baseline by $P5$6, and the abstract and conclusion repeat this as a $P5$7 improvement over conventional feature pyramids (Chen et al., 26 Sep 2025).

The paper also reports an ablation sequence in which HEPAN is added after a $P5$8 detection head. That ablation contains inconsistencies between printed table values and paragraph text: for example, the table formatting and prose give different values for baseline and intermediate configurations. Nonetheless, the paper presents a consistent intended interpretation: HEPAN improves $P5$9 by $640\times640$0 and $640\times640$1 by about $640\times640$2, with modest increases in parameters and FLOPs (Chen et al., 26 Sep 2025).

Qualitative analysis is also used to characterize HEPAN’s effect. Heatmap visualizations comparing YOLOv8-S with HierLight-YOLO-S are described as showing stronger responses on small objects, with “3–5× higher intensity values compared to baseline YOLOv8,” tighter concentration around object boundaries, and improved attention allocation to small objects and their surroundings. The paper interprets these findings as evidence that hierarchical fusion better preserves fine-grained features, reduces missed tiny objects, and sharpens localization (Chen et al., 26 Sep 2025).

These results were obtained in a framework built on Ultralytics YOLOv8, using VisDrone2019, $640\times640$3 inputs, Adam with learning rate $640\times640$4, momentum $640\times640$5, and training for $640\times640$6 epochs on an NVIDIA RTX 6000 Ada. For the full HierLight-YOLO system rather than HEPAN alone, the paper reports $640\times640$7 for small objects $640\times640$8 while maintaining real-time speed of $640\times640$9 FPS. The paper attributes the overall gain to the complete architecture, with HEPAN described as one of the core reasons through improved hierarchical feature fusion (Chen et al., 26 Sep 2025).

6. Relation to PANet and other pyramid necks

HEPAN is explicitly positioned as an extension of the PANet lineage rather than a departure from it. PANet introduced the information-flow perspective on feature pyramids, arguing that FPN propagates deep semantic information downward but does not propagate low-level localization information upward efficiently. Its bottom-up path augmentation shortened the path between low-level detail and top-level features, replacing a route through the deep backbone trunk—described as potentially $P2$0 layers—with a shortcut of fewer than $P2$1 layers (Liu et al., 2018).

Relative to PANet, HEPAN preserves the idea of bidirectional aggregation but changes the design emphasis in several ways. First, it is built around four-level neck inputs from $P2$2 to $P2$3, with $P2$4 treated as a first-class input to the neck rather than only as an optional extension. Second, it adds per-level $P2$5 compression through HFCC before fusion. Third, it introduces dense hierarchical skip connections beyond the adjacent-scale routing that characterizes standard PANet bottom-up augmentation. Fourth, it frames the neck explicitly as an optimization aid through strengthened gradient flow and residual-style intermediate connections. Fifth, it is specialized for anchor-free UAV detection with extremely small targets rather than for proposal-based instance segmentation (Chen et al., 26 Sep 2025); (Liu et al., 2018).

The comparison with PANet and BiFPN in the HierLight-YOLO study indicates that HEPAN is slightly heavier but more accurate as a neck. This makes its role accuracy-oriented rather than efficiency-oriented. The lightweight character of the final detector comes from coupling HEPAN with IRDCB and LDown, not from HEPAN in isolation (Chen et al., 26 Sep 2025).

A common misconception would be to treat HEPAN as merely “PANet with one extra scale.” The paper’s description supports a narrower and more technical distinction: HEPAN is defined not only by inclusion of $P2$6, but by the combination of hierarchical channel compression, bidirectional fusion, dense cross-level coupling, and residual-style gradient-preserving connections (Chen et al., 26 Sep 2025).

7. Limitations, scope, and prospective development

The paper does not present a dedicated “limitations of HEPAN” subsection, but its trade-offs are explicit in the reported results. HEPAN is slightly heavier than PANet and BiFPN, with more parameters and more FLOPs, and therefore is not by itself a lightweight solution. The final efficiency of HierLight-YOLO depends on combining an accuracy-increasing neck with lightweight modules such as IRDCB and LDown that reduce the overall network cost (Chen et al., 26 Sep 2025).

The scope of the evidence is also specific. The most detailed evaluation is on VisDrone2019, a UAV dataset with $P2$7 annotated images, image resolutions from $P2$8 to $P2$9, and $160\times160$0 categories. This supports the claim that HEPAN is effective for UAV small-object detection, but the paper does not provide a separate cross-domain generalization study for HEPAN as an abstract neck design (Chen et al., 26 Sep 2025).

The future directions stated in the paper concern the overall framework rather than HEPAN alone: multi-spectral imaging for all-weather operation, adaptive compression under dynamic resource constraints, and 3D spatial awareness for better localization. This suggests that the architectural logic of HEPAN—preserving fine detail while aggregating semantics across scales—could be extended toward richer sensing modalities or more adaptive computation, but that implication is not developed as a HEPAN-specific proposal in the paper (Chen et al., 26 Sep 2025).

In summary, HEPAN is best understood as the principal feature-fusion innovation in HierLight-YOLO: a UAV-oriented, four-level, bidirectional neck that augments PANet-style path aggregation with hierarchical $160\times160$1 compression and dense cross-level connectivity in order to preserve $160\times160$2-level detail, strengthen gradient flow, and improve small-object detection under severe spatial and semantic imbalance (Chen et al., 26 Sep 2025).