Abstract: At present, China's textile production is huge, and textile import and export trade affects the domestic economic development.Now, the most prominent problem of textiles is quality, of which fabric pilling condition is an extremely important part. The domestic assessment of pilling grade is mainly carried out by professional scientific personnel in some specific scenarios with the standard pilling samples for comparison, for which there are many limitations. First of all, the rating results may be affected by the subjective influence of the testers. Secondly, this method is time-consuming and costly, so there is a need for an objective rating system to grade the fabric pilling.
In recent years, foreign scholars have began to use the computer to process some traditional images by extracting such parameters as the size, number, density, volume and other parameters of the pilling through traditional image processing methods, and such parameters can be used for fabric pilling rating. With the development of deep learning, high accuracy, convenience and other advantages are becoming increasingly prominent and they are widely used by domestic researchers. In the image field, feature extraction by convolutional neural network is effective and avoids the problem of subjective feature extraction. Therefore, we proposed a new Wide-SqueezeNet network for objective rating of fabric pilling based on deep learning.
In this paper, two kinds of woven fabrics with different compositions and contents were used as samples, a ball-box pilling instrument was used to obtain different grades of pilling samples, and the fabrics were put under the light source for image acquisition by using a grayscale camera. A total of 4,376 samples of both kinds were collected. As for the network model, SqueezeNet with fewer parameters and fast training was used as the main body of the network which was innovatively designed. The network model has ten layers, but the number of parameters is small, so the expression ability of complex problems is weak. The new Wide-Fire module was formed by adding a short connection to the original Fire module and using two 3×3 small convolutions to obtain the information of the pilling feature map at different scales and fusing the features with the output of the original Fire module, while using a depth-separable convolution to replace the ordinary convolution in the network to reduce the computation to increase the training speed. A Wide-SqueezeNet network model with deep separable convolution was finally designed.
After the training, the accuracy of Wide-SqueezeNet with improved Fire module is 2% higher than that of the base network, and the accuracy of Wide-SqueezeNet with deep separable convolution is increased by 0.5% and the speed is improved. The final network model is significantly more accurate than some classical classification network models. Two 3×3 convolutional kernels and one 5×5 convolutional kernel are used for training, and the results show that the accuracy of the network with two 3×3 convolutional kernels is higher, so the two 3×3 convolutions are used for feature extraction.
The experimental results show that the improvements in this paper improve the accuracy of the network classification, and the model size and computational effort are basically the same compared to the original network. Finally, the reliability of the network is further verified by the feature map and heat map of the network output, which proves that Wide-SqueezeNet is reliable in the objective rating method of woven fabric pilling. The comprehensive evaluation shows that the network model proposed in this paper can meet the requirements of pilling rating in the fabric industry.

Key words: woven fabrics, pilling, feature fusion, deep separable convolution

摘要： 为了有效克服目前人工检测速度慢、误差大、主观性强的问题，更快速、准确、客观地对机织物起毛起球样本评级，提出了一种多尺度特征融合的Wide-SqueezeNet网络。首先制作了两种成分不同的机织物起毛起球数据集。根据机织物起毛起球图像中毛球形状大小不一以及分布不均的特点，在网络中改进Fire模块，其中增加了短连接来解决训练中梯度发散等问题，在短连接中使用两个3×3小卷积核来减少计算量并且获取不同尺度的特征图信息，增强网络的特征提取能力来提高准确率；其次为了减少整体的计算量，在网络预处理时将图像统一到224×224大小，并且将网络中普通卷积替换为深度可分离卷积。结果表明，通过使用多尺度特征融合和深度可分离卷积来改进网络，机织物起毛起球的客观评级准确率可以达到99.333%。相比于基础网络SqueezeNe、Resnet、MobileNet、DenseNet、ShuffleNet，该方法的提升分别为2.220%、1.777%、2.666%、1.333%和2.220%。与人工检测需要几分钟到十几分钟不等相比，该网络只需要0.072 s即可检测一幅图像，检测速度大大提高。

关键词: 机织物, 起毛起球, 特征融合, 深度可分离卷积