作为一种传统的纺织产品,金属丝网在工业生产、日常生活、科研等领域起着举足轻重的作用,而金属丝网在编制过程中,表面会产生斑点、断线等缺陷,严重影响金属丝网的质量.为保障产品质量,研究了一种基于改进Faster RCNN算法的金属丝网表面缺陷检测方法.首先,为提高模型缺陷特征提取能力,特征提取网络选用深度残差网络(ResNet50)代替原视觉几何群网络(VGG16),并引入注意力模块;随后,训练过程中利用有预热的余弦退火学习率衰减机制,以提高网络检测精度;同时引入k-means算法和遗传算法,设计了更适合金属丝网数据集的锚框尺寸,以提高候选框的精度,解决缺陷定位不准的问题.经实验验证,利用改进Faster RCNN算法检测的平均精度均值(mean average precision,mAP)达86.95%,较原Faster RCNN算法提高18.81%,为金属丝网缺陷的检测提供了一个有效可行的方案.
Defect detection method of wire mesh based on improved Faster RCNN
As a traditional textile product,the wire mesh plays a pivotal role in various fields including industry,daily life,scientific research,etc.In the preparation process of wire mesh,the surface area suffers from various defects such as spots and broken lines,leading to inferior quality of the wire mesh.In this paper,an improved Faster RCNN algorithm was proposed to realize the intelli-gent defects detection of wire mesh.In order to improve the defect feature extraction ability of the model,the deep residual network(ResNet50)was selected to replace the original VGG16 and an attention module was also introduced.In the training process,the preheated learning rate attenuation mechanism of cosine annealing was used to improve the accuracy of network detection.At the same time,k-means algorithm and genetic algorithm were introduced to improve the precision of candidate frame of wire mesh data-set and solve the problem of inaccurate location of defects.With this improved Faster RCNN algorithm,the mean average precision(mAP)of defect detection reaches 86.95% ,which is 18.81% higher than that of the original Faster RCNN,providing an effective and feasible scheme for the detection of wire mesh defects.
NETL
NSTL
万方数据
