Image identification of cashmere and wool fibers based on the improved Xception network
Yaolin Zhu, Huang JiaYI, Yunhong Li, Wenya Li
Abstract
In order to solve the problem of insufficient features and overfitting in network training, an image identification method of cashmere and wool fibers based on an improved Xception network is proposed. Firstly, the normalized fiber image is input into the Xception network to extract the deep features of the fiber image by the convolution layer and the max-pooling layer. Then, an improved Swish activation function is proposed to reduce the overfitting in the whole connection layer. Finally, the Sigmoid classifier is used to classify fiber features. The experimental results show that the identification accuracy of the model is the percent of 98.95 and at least the percent of 2 higher than the original Xception network. It is verified that the improved model can extract more information about fiber features, improving the identification effect of fiber images.