基于多尺度图像增强结合卷积神经网络的纸病识别分类

Paper Defects Classification Based on Multi-scale Image Enhancement Combined with Convolution Neural Network

关键词:  图像增强  卷积神经网络  多尺度形态学梯度  图像分类

Key Words:image enhancement  convolution neural network（CNN）  multi-scale morphological gradient  image classification

摘要:针对造纸工业中传统纸病识别分类依赖于特征描述子和分类器的选择问题，提出一种多尺度图像增强结合卷积神经网络的纸病识别分类方法。该方法利用多尺度形态学梯度增强纸病图像的边缘轮廓信息，突出缺陷梯度特征，然后利用卷积神经网络（CNN）学习纸病图像的特征并分类识别，从而实现纸病的准确识别分类。实验结果表明，该方法对纸病识别分类的结果明显优于HOG+SVM、LBP+SVM以及传统CNN方法，在Caltech101、KTH-TIPS以及本课题的数据集上的分类正确识别率分别达到98.44%、99.23%和99.64%。与现有纸病识别分类方法相比，本课题方法不需针对各种纸病进行缺陷特征提取和特征描述，能快速实现纸病的准确识别分类。

Abstract:In paper industry， traditional approaches of paper defects classification depend on the selection of feature descriptors and the selector.To address the problem， an approach of paper defects classification based on multi-scale image enhancement combined with convolution neural network(CNN） was proposed in this paper-Firstly， multi-scale morphological gradient was computed and used to enhance the edge contour information of paper image， the gradient features of defects in a paper image was highlighted.Then， CNN was applied to learn image features and employed to classify paper defects images.Finally， paper defects classification with a high accuracy was achieved.The experimental results showed that the proposed approach obtained a higher accuracy， 98.44%， 99.23%， and 99.64% on Caltech101， KTH-TIPS， and our datasets， respectively， was significantly superior to the methods of HOG+SVM， LBP+SVM， and traditional CNN.Compared to current methods for paper defects classification， the proposed approach did not require to extract defects features using feature descriptors．