基于边缘感知和小样本学习的多尺度带钢表面缺陷分割方法

A Multiscale Segmentation Method of Strip Steel Surface Defect Images Using Boundary Awareness and Deep Learning on Small Datasets

深度全卷积语义分割网络能够提供像素级带钢表面缺陷检测,对于带钢质量控制具有至关重要的作用。但是这类模型大多无法感知缺陷边缘,而且性能往往严重依赖大量精确标注的标签样本,严重影响其实际应用。为了解决以上困难,提出了一种基于边缘感知和小样本学习的多尺度带钢表面缺陷语义分割网络。该网络由两个级联的子网络组成。第一个子网络首先利用改进的一次性聚合模块和特征金字塔注意力模块构建编码器,提取多层级和多尺度特征并降低训练所需的数据量。然后将一系列全局注意力上采样模块作为解码器实现高级特征指导低级特征复原空间信息,并输出初步预测结果。第二个子网络利用一个浅层U-Net对第一个子网络获得的初步预测结果进行细化并增强边缘检测能力。东北大学热轧带钢表面缺陷数据集上的实验证明了该方法对夹杂、斑点和划伤等表面缺陷自动提取的可行性和有效性。

Fully convolutional networks for semantic segmentation provide pixel-level detection of strip steel surface defects,which plays a crucial role in product quality control of strip steel.However,most of these models suffer from the loss of boundary information,and their performance is often heavily dependent on a large number of labeled samples,which limits the application of the approach.Thereby,a multiscale and boundary-aware network for segmentation of strip steel surface defects on small datasets was proposed in this work.The network consists of two cascaded encoder-decoder subnets.The first subnet employs an encoder built with oneshot aggregation modules and a feature pyramid attention module to extract hierarchical and multiscale features and reduce the dependence of performance on training dataset size.Then,a decoder consisting of global attention up-sample modules exploits high-level feature map to guide low-level features recovering the lost spatial information,and generates preliminary prediction results.Finally,the second subnet further refines the prediction results from the first subnet.Experiments on NEU-Seg defect dataset demonstrate the feasibility and effectiveness of this method for automatic extraction of surface defects such as inclusion,patch,and scratches.