基于多尺度原型残差特征融合的连接器缺陷检测

Connector Defect Detection Based on Multi-Scale Prototype Residual Feature Fusion

摘要: 缺陷检测是工业生产中的重要一环，高效的缺陷检测方法能极大地提高工业流水线的生产效率。近年来，基于机器视觉的检测技术在降本增效方面取得了巨大进展，且能够进行全自动式检测，适用于各类检测场景。在工业连接器缺陷检测任务中，鉴于实际缺陷的外观变化可能很大，即缺陷区域可以有各种大小、形状和数量，针对这些问题本文提出了一种可用于工业连接器缺陷检测的基于多尺度原型残差特征融合的深度神经网络。具体来说，本文基于经典分割网络UNet进行改进，将不同尺度的特征和它们对应的原型残差特征融合，并使用自注意力机制来强化特征。同时，为了缓解实际检测环境中存在的正常样本和缺陷样本的数量不平衡问题，本文通过在正常样本上放置缺陷来扩充缺陷样本。实验表明，本文方法相比于基线方法表现出了更好的缺陷分割检测性能。

Abstract: Defect detection is an important part of industrial production, and efficient defect detection methods can greatly improve the production efficiency of industrial assembly lines. In recent years, machine vision-based detection technology has made significant progress in cost reduction and efficiency improvement. This detection technology can perform fully automatic detection and is suitable for various detection scenarios. In the task of industrial connector defect detection, considering that the appearance of actual defects may vary greatly, that is, the defect region can have various sizes, shapes, and quantities, this paper proposes a deep neural network based on multi-scale prototype residual feature fusion for industrial connector defect detection. Specifically, this article improves on the classical segmentation network UNet by fusing features of different scales with their corresponding prototype residual features and using self-attention mechanism to enhance features. At the same time, to alleviate the imbalance in the number of normal and defective samples in the actual environment, we expand the defect samples by placing defects on normal samples. Experiments have shown that our proposed method exhibits better defect segmentation performance compared to the baseline method.