Fiber classification method based on improved ResNet-50 network

Abstract

At present, there are problems such as the subjective influence of classification results, expensive detection equipment, and long detection time in the common fiber classification technology. In view of these issues, a modified ResNet-50 neural network was proposed, which is based on the pre-training of ImageNet dataset, to classify four animal hair fibers with similar fiber structures. Taking TensorFlow and Keras as the framework, the basic models of ResNet-50 and VGG-16 was used for verification, and then the ResNet-50 network structure and parameters was adjusted by adding Dropout layer and data enhancement strategy. Finally, the adjusted ResNet-50 network was tested with the test datasets, and the confusion matrix was introduced for evaluating each performance index of the network. Compared with original ResNet-50 and VGG-16 networks, the experiment result shows that the adjusted ResNet-50 model has better comprehensive classification performance with an average accuracy of 98.88% and an average F1 score of 98.88%.

Keywords

• fiber classification
• residual neural network
• resnet-50
• vgg-16
• transfer learning