CGEMA-YOLOv8: Efficient Multilevel Feature Fusion Pavement Distress Detection Model Based on Two-Step Image Preprocessing and Multistage Joint Training Enhancement

Timely and accurate detection of pavement distress is crucial for maintenance decisions. Despite advances in detection methods to ensure accuracy while increasing computational speed, there is still some room to improve the balance between accuracy and speed. Moreover, some methods overlook the uniform detection of multitype distresses, which impacts overall pavement quality assessment. To solve these problems, this study proposed an efficient multitype distress detection method using the original lightweight model CGEMA-You Only Look Once (YOLO)v8, integrating multistage joint training (MSJT) and a two-step image preprocessing technique involving adaptive histogram equalization (AHE) and sharpening filtering (SF). CGEMA-YOLOv8 has made three improvements to YOLOv8: replacing the C2f module with C2fGhost; incorporating the CGEMA module for efficient multiscale attention in the neck-head connection; and adding a weighted learning layer and feature fusion path. These improvements aim to improve detection accuracy and efficiency. MSJT enhances the model training efficiency and robustness. The AHE and SF help to capture detailed distress features. The effectiveness of the aforementioned was verified by ablation experiments tested on six types of distress, including cracks, potholes, repair, and ruts. The results of the comparative experiment show that the proposed method outperforms the state-of-the-art methods in accuracy, achieving 84.1% mean average precision and 83.1% F1. Meanwhile, the proposed method requires only 14.02 MB of parameters and achieves an inference speed of 78.5 FPS, which is just 3.22 MB larger than the smallest model and only 5.9 FPS slower than the fastest method. This method represents a significant advance in automated pavement distress detection and provides valuable insight for maintenance operations.