Research And Application Of Convolution Neural Network In Profiled Fiber Recognition

Fiber composition detection is involved in image processing, image analysis, pattern recognition and machine learning and other areas of research. With the development of digital image processing technology and machine learning, the research work on pattern recognition of profiled fiber has made a great progress. However, the computer automatically recognizes the profiled fiber is still a difficult problem. This subject is derived from "Non-negative Matrix Factorization Algorithm Based on Constraints and Its Application in Automatic Fiber Recognition" National Natural Science Foundation (No.61472075). The focus of the research is to use digital image processing technology and machine learning to achieve automatic analysis and recognition of profiled fiber.In this paper, the recognition of profiled fiber is concerned, and a new method for fiber recognition based on convolutional neural networks is proposed. We also describe a seven-layer convolutional neural network for profiled fiber recognition. Experiments show that the proposed method can not only extract the high-level features of the fiber automatically based on a large number of fiber samples with category labels, to avoid the cumbersome and complex artificial feature extraction process of traditional fiber recognition methods, but also can effectively reduce the effect of distorted fiber on recognition efficiency, and raise the accuracy of classification and the recognition rate of 90.875% has been achieved. In order to improve the fiber recognition result and improve the accuracy of fiber identification, the main contents of this paper are as follows:(1) Image preprocessing:In order to avoid unnecessary calculation, improve the computing speed, we often need to convert color image to gray image. In addition, the noise in the fiber image can seriously affect the segmentation and recognition of fiber. How to remove noise effectively is also a key step in fiber image preprocessing. In order to filter the noise as much as possible, and to preserve the useful information of the original image, the wavelet threshold denoising method is used to improve the quality of the fiber image. In order to further highlight the useful information in the image, the OSTU binarization method is adopted.(2) Image segmentation:image segmentation is the key step of image processing and image analysis, where the original image is divided into a number of specific interest areas. After the image segmentation, the target image can be used for subsequent feature extraction and classification. Because of the characteristic of extrusion and adhesion, the fiber image segmentation is very difficult. In order to effectively reduce the effect of fiber extrusion adhesion on segmentation, a level set image segmentation algorithm is adopted in this paper.(3) Feature extraction:For the traditional machine learning algorithm, the extraction of image features need to rely on manual experience, and the characteristics of artificial extraction is often difficult to meet the characteristics of the distinguishability, independence and reliability. Threrfore, the feature selection result will have a greate influence on the the accuracy of classification. How to avoid the artificial extraction of features, make the machine automatically learn the most hidden features of the data, will greatly improve the accuracy of recognition results. In this paper, in order to improve the recognition rate of profiled fiber, the feature extraction method based on convolutional neural network is used to extract the characteristic of fiber.