YOLOv8-Ghost-P2-PIoU2 Model

• YOLOv8-Ghost-P2-PIoU2 is an advanced object detection model that integrates Ghost modules, a high-resolution P2 detection head, and enhanced PIoU2 loss for precise localization.
• It leverages lightweight feature generation and adaptive pruning to reduce computational cost while maintaining high mAP scores, ideal for resource-constrained environments.
• The design improves detection of small and complex targets across domains like aerial imaging, PCB inspection, and vehicle detection with real-time efficiency.

The YOLOv8-Ghost-P2-PIoU2 model is a contemporary evolution of the YOLOv8 object detection framework, targeting high-performance detection of small and complex targets in resource-constrained environments. By integrating Ghost modules for network lightweighting, a P2-level detection head for enhanced sensitivity to small objects, and an improved PIoU-based loss for precise localization, this model exemplifies the synthesis of accuracy and computational efficiency. It is positioned for applications such as aerial imaging, flying object detection, PCB defect analysis, and edge deployments, where stringent demands on both accuracy and model size are present.

1. Architectural Components and Innovations

Ghost Backbone Modules

Ghost modules replace standard convolutional layers with parameter- and FLOPs-efficient operations. The mechanism first derives a set of intrinsic feature maps via regular convolution:

$Y' = X * f'$

with $X$ the input and $f'$ the primary convolution filter. Subsequently, inexpensive linear transformations $\Phi(\cdot)$ expand the features:

$$\text{Output Feature Maps} = \text{Concat}(Y', \Phi(Y'))$$

where the total number of output channels is $n = m \times s$, $m$ being intrinsic feature number, $s$ the ghost factor. This design, detailed in edge-focused and PCB defect detection adaptations, achieves significant parameter and computational reduction while retaining expressivity (Pokhrel et al., 23 May 2025, Pingzhen et al., 23 Jul 2025).

Feature Pyramid with P2 Detection Head

Augmenting the traditional FPN/PAN structure, the P2 branch operates at a higher spatial resolution, enabling improved small object detection. SOD-YOLO and similar models demonstrated that introducing P2-based detection:

• Increases sensitivity to tiny targets via access to shallower, more detailed features
• Employs fusion with upsampled deeper features and attention modules for refined multi-scale representation Enhanced variants further introduce dedicated fusion blocks such as C2f-Faster and adaptive scale feature pyramids (ASFP) to manage parameter growth and retain efficiency (Wang et al., 17 Jul 2025, Shi et al., 17 Dec 2024).

Advanced Loss: PIoU2 and Derivatives

PIoU2 loss references refined intersection-over-union formulations tailored for bounding box regression. While YOLOv8 originally employs CIoU, subsequent improvements include:

• Auxiliary boundaries with adaptive scaling (Inner-IoU), imposing adaptive penalties for fine localization

$$b_l = x_{gt} \pm (w_{gt} \cdot ratio), \quad b_b = y_{gt} \pm (h_{gt} \cdot ratio)$$

• MPDIoU, introducing distance-sensitive penalties via keypoints between predicted and ground-truth boxes
• Combined boundary loss (e.g., Inner-MPDIoU) for enhanced guidance in presence of tiny targets

These modifications, shown to produce substantial gains in mAP$_{0.5:0.95}$ (notably a 10% increase over YOLOv8n for PCB defects), increase the model’s localization fidelity without adverse impact on inference speed (Pingzhen et al., 23 Jul 2025).

2. Model Compression and Adaptive Pruning

Efficient real-time deployment further relies on adaptive pruning strategies. Techniques such as LAMP (Layer-Adaptive Magnitude-based Pruning) determine channel importance:

$$\text{score}(u; W) = \frac{(W[u])^2}{\sum_v (W[v])^2}$$

and iteratively remove low-salience channels. Applied to Ghost-based backbones, such pruning reduces parameter count and GFLOPs to as low as 22% of YOLOv8n’s baseline, with mAP$_{0.5:0.9}$ scores not only preserved but increased. Pruning rates are empirically tuned to avoid over-pruning and degradation (Pingzhen et al., 23 Jul 2025).

3. Empirical Performance and Trade-offs

Detection Accuracy

Empirical results on task-specific datasets illustrate the trade-offs and strengths of the Ghost-P2-PIoU2 integration:

Model Variant	mAP$_{50}$	mAP$_{50:95}$	Parameters (M)	GFLOPs	Inference Time (ms)
YOLOv8-Ghost-P2	0.991	0.896	1.6	8.8	1.5
YOLO-ResNet-18	—	0.976	—	—	—
YOLOv8n (baseline)	—	—	2.96	11.3	0.9–1.0

On real-world visual tasks (e.g., PKLot or PCB test sets), YOLOv8-Ghost-P2-PIoU2 achieves substantial mAP increases (up to 10% for strict IoU) and reduced model complexity (Pokhrel et al., 23 May 2025, Pingzhen et al., 23 Jul 2025).

Efficiency

Ghost integration directly translates to reduced training/inference computational burden, which is critical for mobile or edge scenarios. With tailored pruning, these models further increase throughput (FPS) and decrease latency, enabling batch processing at over 1,200 FPS in pruned regimes (Pingzhen et al., 23 Jul 2025).

Trade-offs

Compared to full-capacity backbones (ResNet, EfficientNet), Ghost-based models offer diminished recall and localization accuracy on extremely challenging IoU thresholds, but the penalty is marginal and often outweighed by vastly superior computational efficiency in practical deployments (Pokhrel et al., 23 May 2025).

4. Application Domains

Aerial and UAV Imaging

Enhanced P2 layers and PIoU variants have proven effective in small object detection problems typical of aerial surveillance, where targets often occupy few pixels and exist in cluttered backgrounds. SOD-YOLO and wavelet-transform-C2f-based variants (cf. (Shi et al., 17 Dec 2024)) further demonstrate that these modifications are robust to orientation and crowding.

PCB Defect Inspection

The model’s ability to capture fine-grained anomalies with high mAP at minimal computational footprint makes it suitable for industrial settings demanding real-time detection of tiny object (defect) classes (Pingzhen et al., 23 Jul 2025).

Parking Lot/Vehicle Presence Detection

The lightweight properties and robust feature extraction enable deployment on embedded devices for real-time parking space management, particularly where partial occlusion and variable lighting challenge conventional detectors (Pokhrel et al., 23 May 2025).

5. Theoretical and Practical Significance

The YOLOv8-Ghost-P2-PIoU2 model demonstrates:

• The feasibility of replacing heavy convolutional operations with structured feature expansion (Ghost modules) without sacrificing detection accuracy.
• That auxiliary, high-resolution detection heads (P2) provide substantial gains for small-object localization, especially when paired with effective multi-scale fusion strategies.
• That enhanced, task-sensitive IoU losses (PIoU2/Inner-MPDIoU) serve as powerful surrogates to classical box regression objectives in settings demanding sub-pixel annotation sensitivity.
• That model pruning, judiciously applied, sustains or improves predictive utility while markedly shrinking deployment cost.
• That, collectively, these strategies represent a principled progression for efficient, real-world object detectors across diverse domains.

6. Trends and Comparative Context

Recent literature positions the YOLOv8-Ghost-P2-PIoU2 approach within a broader trend toward parameter-efficient architectures, dynamic/attention-based feature fusion, and advanced loss design. Compared to DecoupleNet and vanilla YOLOv8, such models consistently offer competitive or superior mAPs at reduced parameterization, situating them as models of choice for edge inference and battery-limited computation (Shi et al., 17 Dec 2024, Wang et al., 17 Jul 2025). Further, this architecture aligns with the movement toward universal detection backbones serving both industrial and aerial or remote-sensing tasks.

7. Considerations for Deployment and Future Work

While empirical evidence validates the combination of Ghost, P2 detection, and advanced PIoU losses for lightweight, accurate detection, continued improvements are anticipated from:

• Further optimization of pruning schedules and dynamic kernel selection
• Integration of more expressive fusion mechanisms (e.g., attention modules as in ASF/ScalSeq)
• Extension to regression tasks for rotated or arbitrarily-shaped bounding boxes in aerial and industrial scenarios
• Systematic benchmarking under extreme edge constraints and for highly-imbalance class distributions

The modular nature of the YOLOv8-Ghost-P2-PIoU2 design admits ongoing research into hybrid or task-specialized backbones, paving the way for universal, high-performance detectors tailored to strict latency and accuracy envelopes.

---

In summary, the YOLOv8-Ghost-P2-PIoU2 model encapsulates the contemporary design paradigm for efficient object detection, combining efficient feature generation, high-resolution detection heads, and advanced bounding box regression losses to address real-world detection requirements without compromising practical deployability.