Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (1): 100-107.DOI: 10.3778/j.issn.1002-8331.2206-0100

• Improvement and Application of YOLO • Previous Articles     Next Articles

Improved Small Target Detection Method of Bearing Defects in YOLOX Network

LI Yadong, MA Xing, MU Chunyang, LI Jiandong   

  1. 1.School of Electrical and Information Engineering, North Minzu University, Yinchuan 750021, China
    2.The Key Laboratory of Intelligent Information and Big Data Processing of Ningxia Province, North Minzu University, Yinchuan 750021, China
    3.College of Mechatronic Engineering, North Minzu University, Yinchuan 750021, China
  • Online:2023-01-01 Published:2023-01-01



  1. 1.北方民族大学 电气信息工程学院,银川 750021 
    2.北方民族大学 宁夏智能信息与大数据处理重点实验室,银川 750021
    3.北方民族大学 机电工程学院,银川 750021

Abstract: Aiming at the problems of high miss detection rate of small targets and insufficient model feature fusion in multi-target case of deep learning model in industrial bearing surface defect detection, a small target defect detection algorithm based on multi-attention feature weighted fusion is proposed based on YOLOX. In the backbone network, a more fine-grained Res2Block module for feature extraction is introduced, and a self-attention mechanism is embedded to increase the regional features of hidden small targets and reduce the missed detection rate. It designs a two-way Pyramid feature fusion network with embedded coordinate attention as a weighting condition to improve the interactive fusion ability of shallow detail features and deep high-level semantic features. In the post-processing stage, the Focal Loss function is introduced to increase the learning of the model on the target of positive samples and further reduce the missed detection rate. The experimental results show that compared with the original YOLOX algorithm, the improved algorithm improves the mAP by 4.04 percentage points on the self-made small train bearing surface defect dataset, which significantly improves the recognition rate of dense small targets.

Key words: bearing surface defect detection, YOLOX, self-attention, feature weighted fusion, coordinate attention

摘要: 针对深度学习模型在工业轴承表面缺陷检测中多目标情形下的小目标漏检率高、模型特征融合不充分的问题,基于YOLOX提出一种多注意力特征加权融合的小目标缺陷检测算法。在骨干网络引入特征提取更加细粒度的Res2Block模块,同时嵌入自注意力机制,增加隐性小目标的区域特征,减少漏检率;设计内嵌坐标注意力并作为加权条件的双路金字塔特征融合网络,提升浅层细节特征和深层高级语义特征的交互融合能力;后处理阶段引入Focal Loss损失函数,增加模型对正样本目标的学习,进一步减少漏检率。实验结果表明,与原YOLOX算法相比,改进算法在自制小型列车轴承表面缺陷数据集上mAP提高了4.04个百分点,对小目标的识别率明显提升。

关键词: 轴承表面缺陷检测, YOLOX, 自注意力, 特征加权融合, 坐标注意力