Department of Electrical and Computer Engineering, TamKang University, New Taipei City 251, Taiwan.
Sensors (Basel). 2020 Aug 5;20(16):4356. doi: 10.3390/s20164356.
This paper presents an improved Convolutional Neural Network (CNN) architecture to recognize surface defects of the Calcium Silicate Board (CSB) using visual image information based on a deep learning approach. The proposed CNN architecture is inspired by the existing SurfNet architecture and is named SurfNetv2, which comprises a feature extraction module and a surface defect recognition module. The output of the system is the recognized defect category on the surface of the CSB. In the collection of the training dataset, we manually captured the defect images presented on the surface of the CSB samples. Then, we divided these defect images into four categories, which are crash, dirty, uneven, and normal. In the training stage, the proposed SurfNetv2 is trained through an end-to-end supervised learning method, so that the CNN model learns how to recognize surface defects of the CSB only through the RGB image information. Experimental results show that the proposed SurfNetv2 outperforms five state-of-the-art methods and achieves a high recognition accuracy of 99.90% and 99.75% in our private CSB dataset and the public Northeastern University (NEU) dataset, respectively. Moreover, the proposed SurfNetv2 model achieves a real-time computing speed of about 199.38 fps when processing images with a resolution of 128 × 128 pixels. Therefore, the proposed CNN model has great potential for real-time automatic surface defect recognition applications.
本文提出了一种改进的卷积神经网络（CNN）架构，用于通过深度学习方法利用视觉图像信息识别硅酸钙板（CSB）的表面缺陷。所提出的 CNN 架构受到现有 SurfNet 架构的启发，并被命名为 SurfNetv2，它包括特征提取模块和表面缺陷识别模块。系统的输出是 CSB 表面上识别出的缺陷类别。在训练数据集的收集过程中，我们手动捕获了 CSB 样本表面呈现的缺陷图像。然后，我们将这些缺陷图像分为四类，分别是崩溃、脏污、凹凸和平整。在训练阶段，所提出的 SurfNetv2 通过端到端的监督学习方法进行训练，使 CNN 模型仅通过 RGB 图像信息学习如何识别 CSB 的表面缺陷。实验结果表明，所提出的 SurfNetv2 优于五种最先进的方法，在我们的私有 CSB 数据集和公共东北大学（NEU）数据集上分别实现了 99.90%和 99.75%的高识别准确率。此外，所提出的 SurfNetv2 模型在处理 128×128 像素分辨率的图像时实现了约 199.38 fps 的实时计算速度。因此，所提出的 CNN 模型在实时自动表面缺陷识别应用方面具有很大的潜力。
