Intelligent cashmere/wool classification with convolutional neural network

Abstract

It is generally believed that there are subtle differences in textures and diameters, between cashmere and wool fibers. Thus, automatically classifying the cashmere/wool fiber images remains a major challenge to the textile industry. In this proposal, we introduced a method that uses Convolutional Neural Networks (CNNs) to identify the two kinds of animal fibers. Specifically, a typical CNN was used to extract image features at first step. Then a region proposal strategy (RPS) was used to localize the fine-grained features from the images. We fine-tuned the CNN model by using the features selected by RPS. Experiments on cashmere/wool image set compared to different models verified the effectiveness of the proposed method for feature extraction.