基于数据增强和改进卷积神经网络的织物纬斜检测

Fabric Skewing Detection Based on Data Augmentation and Improved Convolutional Neural Network

纬斜是织物生产和后整理加工中常见的疵点,光电设备检测纬斜的方法效率低且不精确。为了提升纬斜疵点的检测效率,将神经网络运用到纬斜检测中,结合纬斜特征改进卷积神经网络,提出具有循环训练策略的目标识别网络——纬斜检测网络,并在网络中加入正样本回归和多尺度输入,以提升卷积网络的性能。为了获得充足的纬斜样本数据用于网络训练,提出一种纬斜疵点数据增强方法,通过将纬斜图像公式化生成大量纬斜样本,并采用综合比较实验评估纬斜检测网络性能。结果表明,纬斜检测网络在纬斜检测中表现出色,检测精度达到98%,平均F-score达到0.97,同时使纬斜率的误差控制在±8%以内,检测性能优于其他目标检测模型。与YOLO网络相比,纬斜检测网络在真实纬斜样本检测中性能优异,拥有良好的跨数据集检测性能。

Skewing is a common defect in fabric production and finishing processing.The method of detecting skewing by photoelectric equipment is inefficient and imprecise.To improve the detection efficiency of skewing,the neural network was proposed to be applied to skewing.A new object detection model with a recurring training strategy called the skewing detection network(SDN),was built by an improved convolutional neural network based on skewing features.Positive sample regression and multi-scale input were added to improve the performance of convolutional networks.To acquire sufficient data on skewing samples for network training,a data augmentation method for skewing was developed,which generated a large number of skewing samples by formularizing the oblique images.The performance of SDN was evaluated by comprehensive comparison experiment.The results show that the SDN performed outstandingly with a precision of 98%,an F-score of 0.97,and the error of skewing slope is controlled at±8%,better than other target detection models.Compared to YOLOs,the SDN has excellent performance in real skewing sample detection and has good cross-data set detection performance.