关键词: 实时缺陷检测, 结构重参数化, 共享卷积, 组归一化, 模型轻量化, 损失函数

Abstract: A real-time industrial defect detection method was proposed to address the need for efficient defect detection in modern industrial production. The method was based on structural re-parameterization and shared convolution.It aimed to enhance detection accuracy and inference speed while reducing model complexity. In the feature extraction stage,the LRD-ELAN module was designed using partial channel convolution and structural re-parameterization. Re-parameterization modules and deformable convolution were introduced.These strengthened feature representation and reduced redundant computations. A lightweight GS-Head module was designed for the detection head. Ghost shared convolution was used to reduce computational costs. Group normalization was incorporated to improve stability during small-batch training. A lightweight feature scaling layer was introduced in the bounding box prediction branch to enhance multi-scale object detection capability. The inner-MPDIoU loss function replaced traditional CIoU loss function.This further optimized bounding box prediction performance. Experimental results showed that the proposed model improved mAP50 by 2.46 % over the baseline model on the NEU-DET dataset.Parameter size and computational complexity were reduced by 37.10 % and 33.58 %,respectively. The inference frame rate reached 286.1 FPS,an improvement of 10.82 %,and demonstrated good real-time performance. In generalization tests on the PKU-Market-PCB dataset,the mAP50 was increased by 1.91 %. The model demonstrated strong generalization ability and efficiency. It achieved a good balance between speed,accuracy,and complexity. This highlighted its potential for industrial defect detection.

Key words: real-time defect detection, structural re-parameterization, shared convolution, group normalization, model lightweigh-ting, loss function