Impact of Image Preprocessing and Crack Type Distribution on YOLOv8-Based Road Crack Detection

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Road crack detection is vital for infrastructure maintenance and safety.
• Accurate crack detection relies on robust image analysis and machine learning models.
• Optimizing detection performance requires understanding the influence of data characteristics and preprocessing.

Purpose of the Study:

• To evaluate the impact of image preprocessing techniques on YOLOv8s crack detection.
• To assess the effect of dataset balance on the performance of crack detection models.
• To compare the effectiveness of RGB, grayscale, and binarized images for road crack identification.

Main Methods:

• Utilized YOLOv8s for road crack detection across four datasets: CFD, Crack500, CrackTree200, and CrackVariety.
• Experimented with RGB, five grayscale conversion methods, and binarized images.
• Analyzed detection accuracy in relation to image format and dataset class distribution.

Main Results:

• RGB images consistently yielded the highest detection accuracy, preserving crucial color and texture information.
• Grayscale conversion performance varied by dataset and method; binarization generally reduced accuracy.
• Imbalanced datasets led to biased predictions, whereas the balanced CrackVariety dataset showed more generalized detection.

Conclusions:

• Dataset balance is a more critical factor for YOLOv8s crack detection performance than image preprocessing.
• Future work should address class imbalance through data augmentation and resampling.
• Multi-modal fusion approaches may offer further improvements in road crack detection accuracy.