基于改进RT-DETR的连接器表面缺陷检测算法

Connector Surface defect defection algorithmbased on improved RT-DETR

摘要:

针对现有连接器外观缺陷检测方法存在多尺度缺陷特征提取能力不足、模型复杂度偏高等问题,提出了一种基于改进RT-DETR的连接器表面缺陷检测算法；首先在主干网络中,通过FasterNet与EMA注意力机制的融合,减少算法参数量；其次,使用HS-FPN改进Neck部分,增强模型对细微特征的表达能力；此外,还引入Inner-SIoU损失函数替换原损失函数,提高小目标检测精度并强化模型的泛化能力；最后,在自建连接器缺陷数据集上进行检测试验；结果表明,改进后的算法平均检测精度可达92.5%,参数量相较于基准模型降低了47.2%,性能指标均优于基准模型和主流的YOLO系列模型,FPS达到84,实现了高精度与实时性的平衡,能够满足工业环境下连接器缺陷实时检测的需求；

Abstract:

To address the limitations of existing connector defect detection methods, such as insufficient multi-scale defect feature extraction capability and high model complexity, this paper proposes an improved RT-DETR-based algorithm for connector surface defect detection. First, in the backbone network, the fusion of FasterNet and EMA attention mechanism reduces the number of algorithm parameters. Second, the HS-FPN is employed to enhance the Neck section, improving the model"s ability to express subtle features. Additionally, the original loss function is replaced with the Inner-SIoU loss function, which enhances detection accuracy for small targets and strengthens the model"s generalization capability. Experimental results on a self-built connector defect dataset show that the improved algorithm achieves an average detection precision of 92.4%, with parameter count reduced by 47.2%,respectively,compared to the baseline model. The proposed method outperforms both the baseline model and mainstream YOLO series models in performance metrics. The FPS reaches 81, achieving a balance between high precision and real-time performance, thus meeting the requirements for industrial connector defect detection.