YOLOv8x-HICAUps: Attention-Based PCB Defect Detection

Overview

This project presents YOLOv8x-HICAUps, an enhanced object detection framework for accurate Printed Circuit Board (PCB) defect detection. The model is specifically designed to address the challenges of detecting small-scale defects in densely packed PCB layouts with complex backgrounds.

YOLOv8x-HICAUps integrates a HorNet backbone, Convolutional Block Attention Module (CBAM), and attention-based up-sampling to improve feature extraction, feature fusion, and small-object detection accuracy while maintaining computational efficiency.

---

Key Contributions

• Proposed YOLOv8x-HICAUps framework for PCB defect detection
• HorNet-based backbone for enhanced high-order spatial feature extraction
• CBAM-based channel and spatial attention for suppressing background noise
• Attention-based up-sampling for preserving fine-grained defect features
• Optimized detection head focused exclusively on small defect targets
• High detection accuracy with reduced computational overhead

---

Model Architecture

The YOLOv8x-HICAUps architecture consists of three main components:

Backbone

• HorNet with C3HB modules
• Recursive gated convolutions
• Depth-wise separable convolutions for lightweight design

Neck

• CBAM-enhanced feature fusion
• Attention-based up-sampling to preserve semantic consistency
• Improved multi-scale feature aggregation

Detection Head

• Single small-object detection head
• Eliminates medium and large object heads to prevent feature loss
• Optimized for detecting defects smaller than 10×10 pixels

---

Dataset

The model is evaluated on the HRIPCB dataset, a publicly available PCB defect detection benchmark.

• Total images: 693
• Image resolution: ~2777 × 2188
• Defect categories:
  • Missing hole
  • Open circuit
  • Spur
  • Short
  • Spurious copper
  • Mouse bite
• Train–test split: 9:1

---

Training Configuration

• Framework: PyTorch
• Model: YOLOv8x with custom HICAUps modules
• Epochs: 100
• Optimizer: Adam
• Loss components:
  • Bounding box regression loss
  • Classification loss
  • Distribution Focal Loss (DFL)

---

Evaluation Metrics

• Precision
• Recall
• Mean Average Precision (mAP@0.5)
• Mean Average Precision (mAP@0.5:0.95)

---

Results

The proposed YOLOv8x-HICAUps model achieves superior performance compared to baseline detectors.

• Precision: 0.9786
• Recall: 0.9784
• mAP@0.5: 0.983
• mAP@0.5:0.95: 0.537
• Overall detection accuracy: 98.3%

The model demonstrates strong robustness and generalization across all defect categories, particularly for small and subtle defects such as spur and mouse bite.

---

Technology Stack

• Programming Language: Python
• Deep Learning Framework: PyTorch
• Detection Framework: YOLOv8
• Libraries: NumPy, OpenCV
• Tools: Git, Linux

---

Applications

• Automated Optical Inspection (AOI)
• PCB manufacturing quality control
• Industrial defect detection systems
• Embedded and resource-constrained vision systems

---