摘要
针对传统皮革图像缺陷识别准确率和识别效率不高的问题,提出一种改进多层感知分类器的皮革图像缺陷识别方法。首先,以多层感知分类器作为基础网络模型,对其结构进行优化,并选择适宜的激活函数、分类器和权值与偏置更新方法;然后,搭建一个基于改进多层感知分类器的皮革图像缺陷识别模型;最后,提出一套皮革缺陷图像数据集构建方案,通过滑窗裁剪、样本标注、图像增广等获得4类皮革缺陷图像样本,并将该数据集输入至搭建缺陷识别模型中进行缺陷识别。试验结果表明,本模型对孔洞缺陷、划痕缺陷、针眼缺陷和无缺陷4种故障样本的平均精确率、召回率、准确率和F1值分别为96.97%、96.52%、94.99%和96.14,且本模型进行缺陷识别所用时长仅为3.56 s。相较于经典卷积神经网络VGG16、残差网络ResNet10和支持向量机SVM,本模型对皮革图像不同样本的故障识别准确率更高,识别时间更短。由此说明,本模型能够提升皮革图像缺陷识别准确率和效率,模型性能具备优越性和有效性。
Aiming at the low accuracy and efficiency of traditional leather image defects,a method to identify the multi-layer perception classifier was proposed.First,using the multi-layer perception classifier as the basic network model,optimized its structure,and selected the appropriate activation function,classifier and weight and bias update method.Then built a leather image defect identification model based on improved multi-layer perception classifier.Finally,proposed a set of leather defect image data set construction scheme,obtaining four types of leather defect im⁃age samples through sliding window cutting,sample annotation,image enlargement,and the data set was input to the defect identification model for defect identification.The experimental results show that the average accuracy,recall rate,accuracy and F1 values of hole defect,scratch defect,needle hole defect and no defect are 96.97%,96.52%,94.99%and 96.14 respectively,and the defect identification time of this model is only 3.56 s.Compared with classi⁃cal convolutional neural network VGG16,residual network ResNet10,and support vector machine SVM,this model has higher accuracy and shorter recognition time for different samples of leather images.This shows that this model can improve the accuracy and efficiency of leather image defect recognition,and the model performance has advanta⁃ges and effectiveness.
作者
马静
MA Jing(Shaanxi Institute of Technology,Xi'an 710300,China)
出处
《中国皮革》
CAS
2024年第8期40-46,共7页
China Leather
基金
陕西省“十四五”教育科学规划2023年度课题(SGH23Y3055)
陕西国防工业职业技术学院2022年度重点课题(Gfy22-13)。
关键词
多层感知分类器
皮革图像
图像增广
权值与偏置更新
缺陷识别
multi-layer perception classifier
leather image
image enlargement
weight and bias update
defect recognition
