School of Electrical Information and Engineering, Anhui University of Technology, Ma'anshan 243032, China.
Sensors (Basel). 2020 Sep 1;20(17):4939. doi: 10.3390/s20174939.
To meet the practical needs of detecting various defects on the pointer surface and solve the difficulty of detecting some defects on the pointer surface, this paper proposes a transfer learning and improved Cascade-RCNN deep neural network (TICNET) algorithm for detecting pointer defects. Firstly, the convolutional layers of ResNet-50 are reconstructed by deformable convolution, which enhances the learning of pointer surface defects by feature extraction network. Furthermore, the problems of missing detection caused by internal differences and weak features are effectively solved. Secondly, the idea of online hard example mining (OHEM) is used to improve the Cascade-RCNN detection network, which achieve accurate classification of defects. Finally, based on the fact that common pointer defect dataset and pointer defect dataset established in this paper have the same low-level visual characteristics. The network is pre-trained on the common defect dataset, and weights are transferred to the defect dataset established in this paper, which reduces the training difficulty caused by too few data. The experimental results show that the proposed method achieves a 0.933 detection rate and a 0.873 mean average precision when the threshold of intersection over union is 0.5, and it realizes high precision detection of pointer surface defects.
为了满足检测指针表面各种缺陷的实际需求，并解决一些指针表面缺陷的检测难度，本文提出了一种基于迁移学习和改进的级联 RCNN 深度神经网络（TICNET）的指针缺陷检测算法。首先，通过变形卷积对 ResNet-50 的卷积层进行重构，增强了特征提取网络对指针表面缺陷的学习能力。此外，有效地解决了由于内部差异和特征较弱而导致的漏检问题。其次，采用在线硬例挖掘（OHEM）的思想对级联 RCNN 检测网络进行改进，实现了缺陷的精确分类。最后，基于通用指针缺陷数据集和本文建立的指针缺陷数据集具有相同的低水平视觉特征这一事实，将网络在通用缺陷数据集上进行预训练，并将权重转移到本文建立的缺陷数据集上，从而降低了因数据过少而导致的训练难度。实验结果表明，当交并比阈值为 0.5 时，所提出的方法的检测率达到 0.933，平均精度达到 0.873，实现了指针表面缺陷的高精度检测。
