Improvement of Lightweight Convolutional Neural Network Model Based on YOLO Algorithm and Its Research in Pavement...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Infrastructure Monitoring

Background:

• Existing road pit detection models struggle with imperfect feature extraction and practical equipment limitations.
• Ensuring highway traffic line safety necessitates accurate and efficient pavement defect detection.

Purpose of the Study:

• To propose a lightweight target detection algorithm with enhanced feature extraction for road defect detection.
• To improve the accuracy and reliability of pavement defect detection for road safety.

Main Methods:

• Utilized the YOLO (You Only Look Once) algorithm as the base.
• Implemented Bidirectional Feature Pyramid Network (BIFPN) for multi-scale feature fusion.
• Employed Varifocal Loss to address sample imbalance issues.

Main Results:

• The BV-YOLOv5S model demonstrated improved performance on the PCD1 dataset.
• Achieved a 4.1% higher mAP@.5 compared to YOLOv3-tiny.
• Showcased superior accuracy and reliability over YOLOv5S and B-YOLOv5S models.

Conclusions:

• The proposed BV-YOLOv5S network model offers enhanced performance and reliability for pavement defect detection.
• The algorithm meets the high real-time and flexibility requirements for road safety detection projects.