Investigation of a novel automatic micro image-based method for the recognition of animal fibers based on Wavelet and Markov Random Field

Abstract

In the analysis of fiber recognition, the challenge lies in the texture feature extraction. The main aim of this paper is to present a novel texture feature analysis method based on wavelet multi-scale analysis to fully extract texture features of microscopic images resulting in better recognition of similar animal fibers. Thousands of three kinds of similar fiber images including cashmere, sheep wool and goat hair were captured by the optical microscope and the digital camera. They were pre-processed to obtain the enhanced images with background removed. Then the pretreated fiber images were decomposed by 3-layer wavelet transform, four sub-images under the third-layer wavelet decomposition scale were analyzed by Gauss Markov Random Field (GMRF) model and their model parameters were obtained. Through the difference analysis of different kinds of fibers, two model parameters were selected from each sub-image to generate an 8-dimensional feature vectors, which was used to describe the fiber images. The parameters, which were extracted from 1000 images of each kind of fiber, were copied three times and randomly arranged to generate the final data sets. Finally, the data sets were processed by 10-times cross validation method as the training set and testing set of support vector machine (SVM). Ten different recognition rates could be obtained through the experiment, and the mean value was used as the final recognition accuracy of wool and cashmere fibers. The experimental results indicated that the method had a great recognition rate with 90.07% and the performance was robust. It verifies that the method based on wavelet multi-scale analysis is effective for the recognition of similar fibers.