基于改进DANN网络的织物缺陷检测

Fabric Defect Detection Based on Improved DANN Network

针对传统的织物缺陷检测算法普适性不足的问题,提出一种基于改进DANN网络的织物缺陷检测算法。分析了对抗迁移学习领域的DANN网络存在的仅考虑源域和目标域间特征相似的情况和对于复杂图片提取到的特征能力较差的问题。提出了改进的方法,通过在网络中加入MMD层,可以对提取到的目标域特征赋予不同的权重,并使用ResNet50作为特征提取器。将原DANN网络和改进的MMD-DANN网络在织物缺陷图库中进行了测试并对比了二者的缺陷检测结果。结果表明,改进后网络相比于原网络的准确率平均提高了5%左右,且实时性良好,能满足实际工业需求。

Aiming at the poor general applicability of traditional fabric defect detection algorithm,a fabric defect detection algorithm based on improved DANN network was proposed.Firstly,the shortcomings of the DANN network in the fieldof adversarial transfer learning were analyzed.For example,it only considers the situation of similar features between the source and target domains and can t well exact features of the image with complex information.Next,an improved method was proposed for these deficiencies.By adding the MMD layer to the network,different weights could be assigned to the extracted target domain features,and ResNet50 was used as the feature extractor.Finally,the original DANN network and the improved MMD-DANN network were testedin the fabric defect library,and theirdefect detection results were compared.The test results show that the accuracy of the improved network is about 5%higher than that of the original network,with good timeliness,and it can meet practical industrial needs.