Reverse MPDA (RMPDA) Module

• RMPDA is a neural network module that reverses traditional attention flow to preserve fine-grained details and enhance discrimination in aerial detection.
• It utilizes four parallel convolutional branches with dual attention mechanisms (EECA for channels and EPSA for spatial features) to fuse local and global context.
• Empirical studies show that integrating RMPDA with MPDA and AELAN improves detection accuracy and reduces false positives in challenging small-object scenarios.

Reverse Multi-Scale Progressive Dual Attention (RMPDA) is a neural network architectural module developed to address the preservation of fine-grained shape information and local-global feature complementation in vision tasks that demand accurate discrimination between visually similar small objects, such as birds and drones. It was introduced as a core component of the YOLOBirDrone framework to enhance detection and classification performance in challenging aerial scenarios by refining feature aggregation in the model’s backbone and neck (Kaur et al., 13 Jan 2026).

1. Motivation and Context

The canonical Multi-Scale Progressive Dual Attention (MPDA) module in advanced object detection networks, such as YOLOv9, applies dual attention—spatial and channel—progressively from fine to coarse feature scales, with spatial attention focusing on shallow, high-resolution maps and channel attention targeting deeper, low-resolution ones. However, this purely “forward” direction presents two notable limitations: loss of small-object boundary details at deep stages (where channels are numerous but spatial resolution is low) and insufficient reciprocal enrichment of fine-scale high-resolution features with global semantic priors cultivated at coarser levels. RMPDA was developed explicitly to reverse this attention flow, thereby addressing these drawbacks by restoring edge fidelity and boosting context-driven feature enhancement in fine scales (Kaur et al., 13 Jan 2026).

2. Architectural Composition

RMPDA operates over a tensor input $X \in \mathbb{R}^{H \times W \times C}$, where $H, W$ denote spatial dimensions (typically $20$ or $40$) and $C$ is the channel count (e.g., $256$ or $512$). The module is organized in three main steps:

1. Multi-scale feature extraction: Four distinct receptive-field feature branches operate in parallel, each extracting increasingly large-context features via a specific sequencing of $3\times3$ and $5\times5$ convolutions:

$$\begin{aligned} X_1 &= f_{3\times 3}(X), \ X_2 &= f_{3\times 3}(f_{3\times 3}(X)), \ X_3 &= f_{5\times 5}(f_{3\times 3}(X)), \ X_4 &= f_{5\times 5}(f_{3\times 3}(f_{3\times 3}(X))). \end{aligned}$$

Each branch yields $H, W$0 channels, with total output dimensions maintained.

1. Reverse dual attention:
   • Channel attention (EECA) is applied on the lowest-resolution, highest-channel branches ($H, W$1) via global average pooling, adaptive 1D convolution, and sigmoid gating.
   • Spatial attention (EPSA) is performed on higher-resolution branches ($H, W$2) by stacking channel-wise average and max pooled maps, passing them through a $H, W$3 convolution and sigmoid.

$H, W$4

$H, W$5

1. Fusion and final dual attention: All attended branches are concatenated:

$H, W$6

A further cascade of channel and spatial attention (EECA + EPSA) is applied to $H, W$7:

$H, W$8

where $H, W$9 and $20$0 are channel and spatial attention maps, respectively.

3. Mathematical Formulation

Branch operations and attention mechanisms are formalized in the following equations:

• Multi-scale Branches:

$20$1

• EECA Channel Attention:

$20$2

• EPSA Spatial Attention:

$20$3

• Final Fusion:

$20$4

All $20$5 denote sigmoid nonlinearity. Adaptive kernel size in EECA is given by $20$6, typical $20$7, $20$8. EPSA comprises a $20$9 convolutional spatial mask.

4. Shape Preservation and Feature Enrichment

By reversing the canonical MPDA attention order—first enforcing global channel weighting at deep, low-resolution branches and then applying spatial attention to high-resolution branches—RMPDA enables global semantic information extracted at the deepest network levels to propagate upward. This specifically compensates for the attenuation of fine edges and small-object structure that occurs in standard forward-only refinement. In YOLOBirDrone, qualitative visualizations show that RMPDA yields sharper bird boundary delineation and more accurate object box generation. Empirical ablations demonstrate that, in isolation, RMPDA sustains or marginally improves key metrics (e.g., mAP$40$0: 0.940 → 0.939, mAP$40$1: 0.644 → 0.639; see Table below), and when combined with MPDA and AELAN, elevates overall detection accuracy and reduces false positives (Kaur et al., 13 Jan 2026).

Model	mAP$40$2	mAP$40$3	Accuracy (%)	FP (%)
Baseline	0.940	0.644	81.73	5.04
+RMPDA	0.939	0.639	83.04	4.26
+AELAN+MPDA+RMPDA	0.948	0.668	84.91	3.73

mAP: mean Average Precision; FP: False Positives.

5. Integration in Object Detection Frameworks

RMPDA is situated within the YOLOv9 architecture’s neck/fusion stage, operating downstream of the AELAN and MPDA modules. After AELAN produces deep-scale features, RMPDA processes these maps and feeds its output as input to the PAFPN, where it is upsampled and merged with lateral features. This design enables RMPDA to act as a complementary reverse attention mechanism; it directly ingests and modifies representations produced by MPDA, thus realizing bidirectional context-flow between local detail and global semantics within the grid of fusion blocks in the backbone (Kaur et al., 13 Jan 2026).

6. Practical Implementation and Computational Cost

The practical implementation of RMPDA involves four lightweight convolutional branches, EECA and EPSA attention blocks, and a final joint attention gating. Example pseudocode in PyTorch demonstrates branch-wise channel splitting, parallel convolutions, channel and spatial attention, concatenation, and final context gating. On a representative $40$4 feature map, RMPDA incurs an additional ~0.8 GFLOPs and ~1.2 MB of parameters, a negligible overhead compared to the ~200 MB YOLOv9 backbone, confirming its efficiency for deployment in large-scale detection stacks (Kaur et al., 13 Jan 2026).

7. Empirical Evaluation and Significance

Empirical results in the YOLOBirDrone study demonstrate that RMPDA enhances detection accuracy and reduces false positives, especially for small, visually confounding objects. Integrating RMPDA with MPDA and AELAN achieves detection accuracy up to 84.91%, offering superior trade-offs in mAP and FP rates. A plausible implication is that the complementary structure of RMPDA and MPDA generalizes to other scenarios where reciprocal local-global information flow is necessary to avoid detail loss in deep layers—a recurrent bottleneck in dense-object detection tasks. The module’s lightweight construction mitigates computational cost, supporting integration into real-time or resource-constrained settings (Kaur et al., 13 Jan 2026).

8. Potential Extensions and Limitations

RMPDA’s decomposition—allocating channel attention to deep branches and spatial attention to shallow ones—depends on the channel and spatial coherence of features in a given task. Tuning the number of scales, convolution kernel sizes, and attention block implementation parameters (e.g., EECA kernel size, EPSA configuration) is necessary to optimize performance in new settings. While its impact is most pronounced when combined with complementary modules like MPDA and AELAN, it does not introduce a distinct loss term or regularization and remains bounded by the representational power of the overall architecture into which it is integrated. Its utility has been empirically demonstrated in aerial detection, specifically drone vs. bird discrimination, and its design is likely extensible to other vision domains with analogous structural requirements (Kaur et al., 13 Jan 2026).