A Lightweight Pipeline Defect Detection Method via Structural Reparameterization Technique and Ghost Convolution

To address the inference speed problem of pipeline defect detection, this paper proposes a single-stage lightweight pipeline defect detection method based on the reparameterization technique and ghost convolution. Initially, the baseline model selected for this study is You Only Look Once (YOLO)v5. To further enhance the learning ability without increasing model inference time, the reparameterization convolution (RepConv) is used instead of the
$3 \times 3$
convolution of the original model. Subsequently, the ghost convolution (GhostConv) is utilized to build the ghost module, which substantially reduces the model's complexity to achieve the trade-off between accuracy and speed. Finally, an efficient channel attention (ECA) is employed in the backbone network and feature fusion stages to improve the model's feature extraction and aggregation capability. After actual experimental tests, the proposed improvement method reduces about 30% in parameters, reduces about 34% in floating point operations (FLOPs), and improves about 1% in accuracy.