Research On Detection And Algorithms Of Several Non-woven Fibers Based On Morphological Characteristics

With the continuous development of nonwoven technology and intelligent textile technology,the development and preparation of various fiber-like materials have become a hot spot in the research field.The manufacturing technology and process of these fiber materials have brought a lot of application problems to the detection and identification of fiber material components.Wool is a natural animal fiber in the textile industry.It is mainly processed by needle-punching technology in the non-woven field.It is used for high-grade piano button felt cloth and high-grade gasket and gasket sealing material after heat shrinkage.The good acoustic performance and practical dust removal ability can effectively improve the quality of the piano.In wool products,it often mixes a certain amount of cashmere fiber.The content of cashmere fiber determines the quality and processing technology of the finished product as the main raw material,and the precise control of the mixing ratio of cashmere/wool in the preparation process is also Higher requirements are placed on the detection of process technology processes.However,the current market for cashmere/wool is still dominated by manual,which is lacking an objective,real-time,and effective treatment method.Viscose fiber has excellent moisture absorption performance,easy to entangle in non-woven processing,and polyester fiber has good elasticity and corrosion resistance,but it is not easy to dye,and water absorption performance is not as good as viscose fiber.In the non-woven processing process,the two can be mixed to prepare various sealing functional articles.Because of the different application requirements of the finished products,it is necessary to blend the two fibers with different mixing ratios.The preparation of high-quality polyester/adhesive meshes places higher demands on precise and realtime mixing ratio detection technology.Currently,there is no real-time and effective method for determining the mixing ratio of the two by the visual system.Cotton fiber is commonly used as a plant fiber in nonwoven processing and is used in the preparation of medical and health products by using a spunlaced process in nonwovens.The grade of raw cotton determines the quality of the finished product and affects its processing cost.However,the level of the preparation of cotton in the current non-woven process still based on manual grading by judging the impurities and the smoothness of the cotton.There is a lack of a scientific method to judge the preparation of the raw cotton fiber.Classification efficiency and effectiveness are also hard to guarantee.The detection efficiency and effect of the above-mentioned non-woven fiber materials in the process of identification,classification,and detection in the preparation process are difficult to be quantified.In this paper,according to the application requirements,and to improve the diversity of research,select research samples of different granularity.It includes not only polyester or viscose fibers in chemical fibers that can be observed at large magnifications but also cashmere/wool fibers in natural animal fibers that require moderate magnification to detect morphological differences;The morphological characteristics of raw cotton fibers in natural fibers that need to amplified can be studied.The deep feature representation,detection,and related algorithms of fiber materials for non-woven applications are the main research contents.The deep learning and digital image processing techniques are used to study the feature expression and effective feature enhancement methods based on the depth of image representation.The following three problems are mainly studied: 1.Identification of cashmere/wool fibers in non-woven fabrics;2.Classification of fibers based on visual non-destructive non-woven fabrics,and polyester/viscose mixing ratio determination;3.Quality Detection and Classification of Cotton Fiber.Deep learning technology provides theoretical support and application exploration for the development of non-woven intelligent detection technology.The main research contents and innovation points of this paper are summarized as follows:1.Aiming at resolving the problem of identification cashmere/wool in non-woven fabrics,a manual image enhancement overall technical idea was adopted,and the identification algorithm of cashmere/wool fiber based on Mixture-Level was proposed.The algorithm combines the partial image of fiber image with the whole image to improve the eigenvalues that can be effectively distinguished between the two fibers and enhances the characterization ability of the two fibers by the convolutional network.,obtained a classification accuracy of 92.13%.Based on the Mixture-Level sample enhancement algorithm,the Selective Search algorithm is used to enhance the image from the shape,color,size,and texture.A cashmere/wool identification algorithm based on Region Proposal Strategy is proposed.The algorithm extracts the image features of the middle layer to enhance the comprehensive features of the shape,color,size,and texture of the sample.The accuracy and loss function of the training and verification set of multibatch and multi-round model classification and the evaluation of f1-measure prove that the algorithm has good stability,high classification accuracy,and certain applications feasibility.2 The paper researched the identification of various types of non-woven fibers and the problem of the determination of the ratio of polyester-viscose blends based on computer vision.Aim to classify various types of non-woven fibers takes a large number of samples and the long pretreatment time into account.The division between a plurality of different samples is difficult by artificial design features.To get rid of the dependence of the samples,transform the idea of image enhancement,and automatically extract the feature values of the sample with a deeper network,the classification of the residual fiber to the non-woven fiber is proposed for the first time.This method removes the feature enhancement operation of the sample while omitting the denoising operation of the sample.Taking the improvement of network depth and the deeper network extracts more useful features as the main idea,and at the same time,the classification problem of six kinds of non-woven materials in the deeper residual network has obtained 89.68% classification accuracy.Then,for the determination of the blend ratio of polyester/viscose,the blending ratio of the polyester viscose blended product was determined using a segmentation network based on the divide and conquer idea.A texture-based residual network is proposed for texture differences.The algorithm adds a Max-Pooling layer to the residual unit and adds a feature matrix to the texture enhancement unit.The experimental results show that the common residual network has better recognition accuracy for the blend ratio of polyester/viscose blended fibers.At the same time,the texture-wound residual network proposed in this paper determines the blend ratio of viscose/polyester fiber.When the effectiveness of the algorithm is not significantly reduced,the stability of the algorithm is improved,and the time efficiency of using the residual network algorithm is also improved by about 33.33%.The use of a deep residual network provides a new method for visually based mixing ratio determination of nonwoven fibers,and also provides algorithmic support for the blending of nonwovens over online real-time detection.3.For the quality grading problem of raw cotton preparation,the impurity edge detection of the four convolution operators of Sobel,Roberts,LoG,and Canny used as the entry point.The impurities can be found and located by analyzing the convolution operation.The deep convolution networks can use for raw cotton characterization and feature presentation.After comparing with Google Net's Inception-V3 model using migration learning,find that quality of the cotton preparation can directly be used to classify by a deeper residual network.The experimental results show that the method is suitable for the quality of cotton preparation prediction.The classification is efficient and stable,achieving a verification accuracy of 96.97%.In summary,this paper mainly studies the mix-level network and Region Proposal Strategy Network with local promotion as the main idea to identify cashmere wool.At the same time,the determination of the blending ratio of polyester/viscose commonly used in nonwovens by Texture Wise Residual Network with integral segmentation as the mathematical basis to improve the network depth and enhance the texture characteristics also studied.Finally,on the quality grading of raw cotton preparation,compared to the Inception-V3 network through migration learning the residual network is a better choice to solve the problem of quality quantification of raw cotton preparation.