Abstract: Forging surface cracks are harmful with low detection efficiency.To solve the problems existing in the traditional artificial magnetic particle inspection of forging crack scene,a forging surface crack real-time detection algorithm was proposed.The magnetic particle inspection images from the flaw detection workshop in the heavy truck steering knuckle production line were collected,and a forging surface crack dataset was produced.A lightweight multi-scale convolution module named LMSConv was proposed to achieve single-module cross-scale feature extraction.Large Separable Kernel Attention Spatial Pyramid Pooling-Fast(LSKA-SPPF)model was  proposed to further enhance the overall cross-scale feature fusion capability.The Bbox loss function was changed to Focal EIoU to improve the anchor box regression accuracy.Experimental results on forging surface crack dataset and NEU-DET dataset showed that the proposed algorithm had high detection accuracy and low complexity.Compared with other mainstream one-stage object detection algorithms,the proposed algorithm reduced the missed detection rate and false detection rate with high robustness,which could fulfill the needs of forgings cracks inspection.

Key words: surface defect detection, cross-scale feature, forging cracks, one-stage object detection

摘要： 锻件表面裂纹危害性大,检测效率低,为解决传统人工磁粉检测锻件裂纹场景中存在的问题,提出一种锻件表面裂纹实时检测算法。首先采集重卡转向节生产流水线中探伤车间的磁粉检测图像,制作锻件表面裂纹数据集;然后提出轻量多尺度卷积模块LMSConv,实现单模块跨尺度特征提取,同时提出大型可分离内核注意力快速空间金字塔池化模块LSKA-SPPF,进一步加强整体跨尺度特征融合能力;最后,更改Bbox损失函数为Focal EIoU,以提高锚框回归精度。锻件表面裂纹数据集与NEU-DET数据集上的实验结果表明算法检测精度高,复杂度低,对比其他主流单阶段目标检测算法,所提算法漏检率和误检率降低,具有较强的鲁棒性,能够满足锻件裂纹检测的需要。

关键词: 表面缺陷检测, 跨尺度特征, 锻件裂纹, 单阶段目标检测