# RGS-YOLO: A lightweight solution for surface defect detection in wind turbines

## TL;DR

This paper introduces RGS-YOLO, a lightweight model for detecting defects on wind turbines, improving detection accuracy while reducing computational costs.

## Contribution

The novel integration of RFAConv, GSConv, and SCConv modules creates a lightweight yet effective defect detection model for wind turbines.

## Key findings

- RGS-YOLO reduces model size by 1.16 MB and floating-point operations by 3.5 G compared to YOLOv8n.
- The model achieves a 3.7% improvement in mean Average Precision (mAP) over YOLOv8n.
- The model maintains high accuracy while being lightweight and computationally efficient.

## Abstract

The defects in wind turbines not only affect energy generation efficiency but can also lead to significant damage if not repaired promptly. To address the challenges of low detection efficiency and high costs in the real-world industrial scenario of wind turbine defect detection, we have designed a lightweight detection model. First, this study introduces Receptive-Field Attention Convolution(RFAConv) and develops the Cross Stage Partial with 2 convolutions and feature fusion-Receptive-Field Attention Convolution(C2f-RFAConv) module, integrating it into the backbone network. This approach allows the model to focus on spatial features while accurately capturing local information in each region through its receptive field, significantly enhancing its feature extraction capabilities. Additionally, we incorporate Group Shuffle Convolution(GSConv) in the neck network to ensure that the model remains lightweight while maintaining a high level of accuracy. In the design of the detection head, we leverage the low redundant computation capability of Spatial and Channel reconstruction Convolution(SCConv), along with its ability to promote the learning of representative features, to develop a detection head-SCConv Head-that integrates classification and detection with low computational cost and parameters. All experimental results are reported as the average of no fewer than three independent runs to ensure the stability and reliability of the results. Experimental results show that, compared to the original You Only Look Once version 8 nano(YOLOv8n), our model reduces its size by 1.16 MB and decreases the floating-point operations by 3.5 G while improving the mean Average Precision (mAP) by 3.7%. These results demonstrate the effectiveness of our model in achieving lightweight performance.