摘要
针对布匹瑕疵数据集分辨率高、瑕疵区域小,使用已有图像分类算法识别效果差的问题,提出了一种基于经典目标检测算法Fast RCNN改进的图像识别算法,用目标检测的流程完成图像识别任务。算法利用布匹瑕疵区域小的特性,在图像中生成大量潜在的瑕疵区域,利用卷积神经网络预测潜在区域包含瑕疵的概率,在算法的后期对潜在区域的概率进行合并,最终达到识别布匹中存在瑕疵的概率。在一个具有3331张高分辨率图片的数据集上进行实验,结果表明,本文算法比OurNet以及已有的图像分类算法具有更好的性能。
Fabric defect dataset usually has high resolution but small defect area,which is different from common image classification datasets(such as ImageNet,etc).When applying exist classification algorithms to fabric defect dataset directly,it cannot achieve expected accuracy.To address this problem,this paper proposes a new classification algorithm based on improved Fast RCNN.For an image with small defect area,we follow the pipeline of object detection,generating lots of ROI(Region of Interest),extracting feature map with deep convolutional neural network and predicting each ROI′s class probability.On the last stage,we combine all ROIs′class probabilities to get the full image′s class probability.Experiments performed in a fabric defect dataset which has 3331 high resolution images show that our algorithm outperforms OurNet and exist classification algorithm.
作者
车翔玖
刘华罗
邵庆彬
CHE Xiang-jiu;LIU Hua-luo;SHAO Qing-bin(College of Computer Science and Technology,Jilin University,Changchun 130012,China)
出处
《吉林大学学报（工学版）》
EI
CAS
CSCD
北大核心
2019年第6期2038-2044,共7页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金项目（61672260）
关键词
计算机应用
布匹瑕疵识别
卷积神经网络
图像识别
目标检测
computer application
fabric defect recognition
convolutional neural network
image classification
object detection
