Abstract:

In order to solve the problem of the difficult in identifying of cashmere and wool fiber images, a novel identification method based on convolutional neuron (CNN) and the deep learning theory was proposed. A sigmoid classier was used to carry out coarse classification on fiber image features extracted by the CNN. The reaults were verified according to the validation datasets, and the optimal weights of the network was recorded. Part-level features of each sample image are extracted by an improved part-level augment object-level network based on previously obtained parameters. In addition, the part-level and object-level features were fused, and a new network model was established based on the fused features. On this basis, Ordos’s standard cashmere and wool dataset was used to train the network for 50 times, and the best accuracy is 92.1%. The experiments results demonstrate that the method for classifying cashmere and wool images, based on the CNN features, can be applied to cashmere and wool or discriminate similar fibers identification successfully. However, a large number of sample sets and validations are required for commerical use.

Key words: cashmere, wool, image identification, convolutional neuron networks