Abstract: The magnetic tile image has the characteristics of uneven illumination,.complex surface texture and low contrast,but its defects are hard segmented with the traditional defect detection algorithms.For this reason,a defect detection algorithm based on Mask Region-based Convolutional Network (Mask R-CNN) was proposed.The image was preprocessed by contrast limited adaptive histogram equalization.Then Residual Network 50 (ResNet50) was used to construct the Feature Pyramid Network (FPN) to acquire image information and extract features.Then,the region of interest of the defect region were extracted by Region Proposal Network (RPN) to obtain the corresponding anchor frame,and the pixel class inside the region of interest was predicted by Fully Convolutional Network (FCN) to realize the defect.Through the fully connected layer of the network,the prediction the category and the corresponding anchor frames of each interest region was realized.The experimental results showed that the proposed algorithm had strong generalization ability,and strong robustness.It could accurately segment the surface of the magnetic tile image with complex texture,uneven illumination and low contrast.

Key words: surface defect detection, mask region-based convolutional network, feature pyramid network, magnetic tile

摘要： 磁瓦图像具有光照不均、表面纹理复杂、对比度低等特点,针对传统的缺陷检测算法难以准确分割其中缺陷的问题,提出基于掩膜区域卷积网络(Mask R-CNN)的缺陷检测算法。该算法首先通过限制对比度的自适应直方图均衡化方法对图像进行预处理;然后,采用残差网络50(ResNet50)构建特征金字塔网络(FPN)获取图像信息并提取特征,再采用区域建议网络(RPN)提取缺陷区域的感兴趣区域,得到相应的锚框,并通过全卷积神经网络(FCN)对感兴趣区域内部的像素类别进行预测,以实现缺陷分割;最后通过网络的全连接层实现每个感兴趣区域所属类别和相应锚框坐标的预测。实验结果表明,该算法具有较强的泛化能力,可以对表面存在大量纹理复杂、光照不均和对比度低的磁瓦图像进行精确的缺陷分割,具有较强的鲁棒性。

关键词: 表面缺陷检测, 掩膜区域卷积网络, 特征金字塔网络, 磁瓦