计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (3): 270-279.DOI: 10.3778/j.issn.1002-8331.2209-0316

• 图形图像处理 • 上一篇    下一篇

基于改进特征金字塔的轮胎X光图像缺陷检测

吴则举,宋丽君,冀杨   

  1. 青岛理工大学 信息与控制工程学院,山东 青岛 266520
  • 出版日期:2024-02-01 发布日期:2024-02-01

Tire X-Ray Image Defect Detection Based on Improved Feature Pyramid Network

WU Zeju, SONG Lijun, JI Yang   

  1. College of Information and Control Engineering, Qingdao University of Technology, Qingdao, Shandong 266520, China
  • Online:2024-02-01 Published:2024-02-01

摘要: 针对轮胎X光缺陷图像背景纹理复杂、缺陷目标较小的问题,在带有特征金字塔结构的Faster R-CNN网络基础上做出改进,提出一种轮胎缺陷检测网络。针对特征金字塔底层感受野较小的问题,将空洞-深度可分离卷积和多分支结构结合,设计了轻量型的感受野扩增模块。针对特征金字塔顶层特征图分辨率不足的问题,提出一种新型金字塔结构BE-FPN(bottom embedding feature pyramid network),将底层特征图经过感受野扩增模块后与顶层信息快速归一化融合,减少了底层特征传递到顶层特征时过长路径导致的信息损失。此外,在ResNet-50网络引入动态可学习的激活函数Meta-ACON,有效改进了网络检测性能。改进后的网络取得了94.07%的平均精度均值(mAP),与基线网络相比,精度提升4.5个百分点,正判率提高了14.4个百分点,参数量仅增加4.77%,每张图片检测时间仅增加0.009?s。实验结果表明,设计的网络在轮胎X光缺陷检测方面具有优良的性能,能够满足工厂生产线对检测精度和检测时间的要求。

关键词: 深度学习, 缺陷检测, 特征融合, Meta-ACON

Abstract: To solve the problem of complex background texture and small defect target of tire X-ray defect images, a tire defect detection network is proposed based on the Faster R-CNN network with feature pyramid network. Aiming at the problem that the receptive field at the bottom of the feature pyramid is small, a lightweight receptive field amplification module is designed by combining the dilated-depthwise separable convolution and the multi-branch structure. For the problem of insufficient resolution of the top feature map in the feature pyramid, a new pyramid structure BE-FPN (bottom embedding feature pyramid network) is proposed. The bottom feature map is fast normalized fused with the top layer after through the receptive field amplification module, which reduces the information loss caused by too long path when the bottom feature is transferred to the top feature. In addition, the introduction of a dynamic and learnable activation function Meta-ACON in the ResNet-50 network effectively improves the network detection performance. The improved network achieves 94.07% mean average precision (mAP), which is 4.5?percentage points higher than the baseline.The rate of correct judgments is increased by 14.4?percentage pionts, the parameter amount is increased by 4.77%, and the detection time of each image is only increased by 0.009?s. The experimental results show that the network designed in this paper has excellent performance in tire X-ray defect detection, and can meet the requirements of factory production line for detection accuracy and detection time.

Key words: deep learning, defect detection, feature fusion, Meta-ACON