PV-YOLO: Collaborative Lightweight Improved YOLOv8 for Micro-Defect Detection of Photovoltaic Cells in UAV Edge Inspection

Technical Summary: PV-YOLO for Micro-Defect Detection in Photovoltaic Cells

Problem Statement
The reliable operation of photovoltaic (PV) modules is critical for energy efficiency, yet micro-defects such as micro-cracks, hotspots, and broken grids often go undetected by traditional manual inspection or standard machine vision methods. These defects, often subtle with low contrast against complex backgrounds, lead to significant power degradation and financial losses. While deep learning has advanced defect detection, existing models face a critical trade-off: high-accuracy models are often too computationally heavy for real-time edge deployment (e.g., on UAVs), while lightweight models frequently suffer from insufficient sensitivity to small, low-contrast targets and high rates of missed detections. Furthermore, static multi-scale feature fusion strategies often fail to adapt to the diverse sizes of PV defects.

Methodology
The authors propose PV-YOLO, a collaborative lightweight improved YOLOv8n model designed specifically for UAV edge inspection. The architecture integrates four custom modules to balance detection precision with computational efficiency:

1. C2f_faster Backbone: To optimize gradient transmission and reduce redundant computation, the standard C2f module's Bottleneck blocks are replaced with FasterBlock modules. These utilize Partial Convolution (PConv) to extract spatial features from a subset of channels while leaving others unchanged, significantly lowering floating-point operations (FLOPs) and memory access volume.
2. GhostConv: To address feature redundancy, standard convolutions are replaced with GhostConv. This module generates a subset of intrinsic feature maps via standard convolution and then produces additional "ghost" feature maps using low-cost linear operations, effectively compressing the model without losing key defect information.
3. Efficient Multiscale Attention (EMA): To enhance sensitivity to faint micro-defects, the EMA module is introduced. It employs feature grouping, parallel subnets, and cross-domain learning to fuse local and global attention features. This allows the model to focus on subtle grayscale variations and edge information, distinguishing true defects from background noise.
4. Adaptive Feature Pyramid Network (AFPN): Replacing static fusion strategies, the AFPN dynamically adjusts the fusion weights of multi-scale features based on defect size information. It incorporates direct feature fusion between non-adjacent levels and adaptive spatial fusion to strengthen the representation of small and medium-sized targets.

The model was trained and evaluated on the PVEL-AD dataset, comprising 4,500 near-infrared (EL and IR) images with 12 distinct defect categories. To address class imbalance, targeted data augmentation (geometric and photometric transformations) was applied to minority classes.

Key Contributions

• Optimized Backbone: The C2f_faster module reduces backbone FLOPs by 21.0% compared to the baseline YOLOv8n while enhancing gradient flow for low-contrast features.
• Lightweight Convolution: The integration of GhostConv reduces the model's parameter volume by 10.96% and FLOPs by 3.7% compared to the baseline.
• Enhanced Sensitivity: The EMA module specifically targets the low sensitivity of general attention mechanisms to subtle PV defects, increasing the recall rate for micro-cracks by 7.2% over the baseline.
• Dynamic Fusion: The AFPN module improves the mAP@0.5 for small defect detection by 3.14% compared to traditional static FPNs by adapting to varying defect sizes.
• Edge-Deployable Solution: The complete system is validated on NVIDIA Jetson Orin hardware with CUDA and TensorRT acceleration, achieving real-time inference speeds.

Experimental Results
Experiments conducted on the PVEL-AD dataset demonstrate that PV-YOLO achieves a 72.3% mAP@0.5, which is 3.13% higher than the original YOLOv8n baseline. Notably, the recall for micro-cracks increased by 7.2%. Despite these accuracy improvements, the model maintains a lightweight profile with 2.68M parameters and 7.8 GFLOPs, representing a 2.46% reduction in FLOPs compared to the baseline.

In terms of deployment, the model runs at 140 FPS on the NVIDIA Jetson Orin edge device. Comparative analysis against other state-of-the-art lightweight models (e.g., YOLO11n, YOLO13n) and heavier architectures (e.g., Faster R-CNN, YOLOv8x) confirms that PV-YOLO offers a superior balance of accuracy and efficiency, outperforming newer lightweight variants in mAP and recall while maintaining significantly lower computational costs.

Significance
The paper claims that PV-YOLO provides a practically deployable solution for the real-time, automated inspection of photovoltaic cells in industrial scenarios. By successfully addressing the limitations of gradient flow, feature redundancy, and static fusion in existing lightweight detectors, the proposed method enables UAVs to perform efficient, large-scale inspections with high precision. This capability is crucial for minimizing power degradation losses and reducing the labor and time costs associated with traditional manual inspection methods. The study validates that deep learning models can be effectively optimized for resource-constrained edge environments without sacrificing the detection of critical micro-defects.