YOLOv7-Rep: a re-parameterization method for surface defect detection in workpieces

The main goal of workpiece surface defect detection is to improve accuracy and real-time performance. Generally, the practice of using a network model with stronger representation ability and performing lightweight design at the same time cannot solve the problem of better performance balance between model training and inference phases. To address this, this paper optimizes and proposes a structural re-parameterization defect detection method: YOLOv7-Rep. First, we proposed a re-parameterization convolutional module: RepC2, which is a new block with cross-stage partial (CSP) connection and efficient gradient flow branch structure. Second, this article integrates the RepC2 module based on YOLOv7 and incorporates the coordinate attention (CA) mechanism in the backbone. Finally, we introduce the WIoUv1 bounding-box regression loss function. Experimental analysis demonstrates that YOLOv7-Rep outperforms object detectors with the same parameter count in terms of both detection accuracy and speed. It achieves detection accuracies of 78.3%, 87.3%, and 83.5% on three industrial component datasets (NEU-DET, TCAP-DET, and GC10-DET), respectively. Compared to YOLOv7, it significantly improves detection frame rates by 4.9% (an increase of 11.3 FPS), while achieving a better performance balance between training and inference phases in surface defect detection tasks. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024.