Surface defect detection for nylon yarn package based on improved VGG model

Abstract

At present, there is still seldom research on how to use neural networks for the surface defect detection of nylon yarn packages. The original VGG has some shortcomings, if there are many network layers, it is difficult to train, and if there are few network layers, it is impossible to learn abundant features to meet the requirements of industrial production. In this paper, an improved VGG-based method was proposed for detecting surface defects on nylon yarn packages. A two-path network structure with 16 and 19 convolutional layers is designed to fuse the features learned by the earlier layers and the fused features will be inputted to the classifier to obtain the final output for defect category. To deal with the small sample size problem which affect to train the neural network effectively, we use data augmentation to process the photos of the input and transfer learning to initialize the model parameters. Our experiments demonstrate that the proposed method improves the accuracy by 1.07% over the VGG16-BN.