Abstract: Aiming at the problem of low detection accuracy of aluminum profile surface defects due to onlyconsidering texture features,a surface defect detection model driven by master-slave feature fusion was proposed.The model construction mainly included three parts.The UNet model optimized by the Focal-Loss loss function was used to complete the segmentation and positioning of samples with uneven defect distribution;then the Convolutional Neural Network (CNN) and the Back Propagation Neural Network (BPNN) was integrated to construct a master-slave feature pre-classification layer that fused image texture features,gradient information and defect shape features;last the final classification of defects was completed by cascading a fuzzy neural network with specific fuzzy rules.The five types of defect samples in the aluminum profile data set of the Ali Tianchi competition were used to experimentally verify the model.The average classification detection accuracy reached 97.2%,which provided a new method for aluminum profile surface defect detection.

Key words: surface defect detection of aluminum profiles, master-slave feature fusion drive, convolutional neural network, back propagation neural network, fuzzy neural network

摘要： 针对因仅考虑纹理特征而造成铝型材表面缺陷检测精度较低的问题,提出一种主从特征融合驱动的表面缺陷检测模型。该模型的构建主要包括3个部分:首先采用经Focal-Loss损失函数优化的UNet模型完成缺陷分布不均匀的样本分割与定位;然后集合卷积神经网络(CNN)与反向传播神经网络(BPNN)构建融合图像纹理特征、梯度信息和缺陷形状特征的主从特征预分类层;最后通过级联特定模糊规则的模糊神经网络完成缺陷的最终分类。利用阿里天池比赛的铝型材数据集中的5类缺陷样本对模型进行了实验验证,平均分类检测精度达到97.2%,为铝型材表面缺陷检测提供了新方法。

关键词: 铝型材表面缺陷检测, 主从特征融合驱动, 卷积神经网络, 反向传播神经网络, 模糊神经网络