Evaluation method of fabric pilling grades based on saliency-based deep convolutional network

Abstract

In order to improve the objectivity of fabric pilling evaluation, a saliency deep convolutional network method for fabric pilling evaluation is proposed. First of all, the fabric pilling instrument is used to generate pilling fabric samples as a nonstandard dataset that is added to the standard fabric pilling dataset. The dataset is expanded through data augmentation to increase the number and diversity of pilling data. Then, a saliency preprocessing model is constructed to achieve the preprocessing of the fabric pilling image dataset by fusing the local and global saliency map. Finally, improvements to the ResNet 34 network model are made. The convolutional layer is improved to achieve small target pilling features enhancement. The residual module in the residual network is improved by using ReLU6 as the activation function, giving a down-sampling convolution on the shortcut branch of each residual block and adding average pooling, which avoids the loss of weight information. An improved attention mechanism module is added to extract fully and learn fabric pilling features according to the channel attention mechanism in parallel with the spatial attention mechanism. The recommended method uses standard and nonstandard pilling fabric samples to expand the number and diversity of the dataset. The improved ResNet 34 network model improves the ability of feature extraction and learning, thus improving the accuracy of pilling evaluation. The experimental results show that the average accuracy of the proposed method is 93.88%, which indicates that the pilling grade evaluation algorithm used can effectively achieve the grade evaluation of fabric pilling.