Improved algorithm for lightweight identification of 3D GPR images of hidden road defects

To address the problems of small real data scale, difficulty in manual interpretation, and insufficient accuracy and efficiency of traditional algorithms in real-time detection of hidden road defects using three-dimensional ground penetrating radar (3D GPR), a lightweight improved algorithm based on the YOLOv8 algorithm YOLOv8-CES was proposed. Based on the collected 3D GPR images of hidden road defects, defect information was annotated and classified to establish a dataset. A Cut_SimAM attention mechanism was proposed based on the parameter-free SimAM attention mechanism to improve small target detection, and it was added to the backbone network to enhance focus on target regions and improve small object detection ability. Based on the C2f structure, a C2f_EFAttention feature extraction module was introduced into the neck network to optimize the feature fusion process, reduce the number of parameters, and improve detection efficiency. The Slide Loss sliding weighted loss function was used in combination with the D-IoU bounding box loss function to accelerate model convergence and improve detection accuracy for difficult samples. Ablation experiments were conducted to verify the effect of each module on model performance, using mean average precision (mAP), number of parameters, floating point operations per second (FLOPs), and frames per second (FPS) as indicators. The detection accuracy and efficiency of the improved algorithm compared with other algorithms were evaluated through comparative experiments. Test results show that in the collected 3D GPR image dataset of defects, the mAP of YOLOv8-CES reaches 61.5%, increasing by 3.6% compared with the baseline model. The number of parameters reduces from 3.0×10⁶ to 2.5×10⁶, and FLOPs reduces from 8.1 GFLOPs to 7.1 GFLOPs; FPS increases by 16.7%. The improved model achieves higher accuracy and lower computational demand in the recognition and classification of hidden road defects. The high accuracy and lightweight design of the YOLOv8-CES algorithm make it more suitable for embedding in 3D GPR detection devices and achieving real-time detection, indicating its potential application value in road detection.