基于改进卷积神经网络的化纤丝饼表面缺陷识别

针对传统人工检测化纤丝饼表面缺陷方法的不足,提出改进的卷积神经网络对正常以及3种常见缺陷丝饼进行分类识别。首先对采集的丝饼图像进行分块处理,然后利用改进的卷积神经网络进行特征提取,采用全局最大池化层代替全连接层,增强了图像对空间变换的鲁棒性,减少了模型参数,并利用softmax分类器进行分类。最后在网络学习过程中提出主动学习方法,用少量标注样本对网络进行训练,选出对提升网络性能最具价值的样本进行标注并加入到训练样本中进行训练检测。结果表明,该方法可有效实现丝饼的缺陷识别,识别准确率达到97.1%,并有效减少了网络所需的标注样本数量,节省大量的标注成本,具有一定的通用性。

Surface defect recognition of chemical fiber yarn packages based on improved convolutional neural network

Focusing on the disadvantages of traditional manual method for defect detection of chemical fiber yarn packages, an improved convolutional neural network was proposed to classify and recognize the normal and three common defective yarn packages. The images of collected yarn package were processed into blocks before features were extracted by using the improved convolutional neural network. The global maximum pooling layer was used instead of the full connection layer, enhancing the robustness of images to spatial translations and reducing the model parameters. Softmax classifier was employed for classification. As a result, an active learning method was proposed for network learning. Firstly, a small number of labeled samples were used to train the network, and then the most valuable samples for improving network performance were selected and labeled, which were then added to the training samples. The experimental results show that this method can effectively facilitate the defect recognition of yarn packages, achieving a recognition accuracy of 97.1%. This method effectively reduces the number of labeled samples required by the network and saves a lot of labeling costs, with a certain degree of universality.