Lightweight Improvement of a Road Defect Detection Model Based on YOLOv8
TL;DR
The lightweight model reduces model complexity but also exhibits a non-negligible decrease in detection accuracy, demonstrating an explicit accuracy-complexity trade-off rather than accuracy-preserving compression.
Abstract
Road-defect inspection based on vehicle-mounted optical imaging requires the image-analysis model to operate under limited on-board storage and computation. This study develops a lightweight target detector for the computational interpretation stage of a visible-spectrum pavement sensing system. In the actual implementation, the original YOLOv8s backbone is replaced with a ShuffleNetV2 feature extractor, while the SPPF module, multi-scale feature-fusion neck, and detection head are retained. The ShuffleNetV2 backbone uses channel splitting, depthwise convolution, pointwise convolution, and channel shuffle to reduce redundant computation. The model is trained for 200 epochs on a seven-class road-defect dataset containing transverse patches, longitudinal patches, transverse cracks, longitudinal cracks, alligator cracks, manhole covers, and potholes. Code-level reconstruction gives 7,327,393 parameters for the improved seven-class model, and the saved best checkpoint is approximately 15 MB. The best recorded validation performance is mAP@0.5 = 0.533 and mAP@0.5:0.95 = 0.340. Compared with the original YOLOv8s result reported in this study, the lightweight model reduces model complexity but also exhibits a non-negligible decrease in detection accuracy. The results therefore demonstrate an explicit accuracy-complexity trade-off rather than accuracy-preserving compression.