Research Of Recognition And Positioning For Heterogeneous Cotton Fiber Image Based On RBF GMDH-type Neural Network

In the spinning process, the traditional method of manual sorting of cotton fiber is difficult,costly and inefficient, which greatly reducing the competitiveness of cotton spinning enterprises.Although there are many ways in the use of machine vision instead of manual sorting at homeand abroad, the disadvantages of poor positioning accuracy, high cost, poor universality, highenvironment demanding and difficult to implement in many fields are ubiquity. With the rapiddevelopment of digital image processing technology, the increasingly widespread application ofphotonics and the maturity of machine vision, it’s entirely feasible to realize precise recognitionand positioning of cotton heterogeneity fiber.With digital image processing technology, this paper designs and implements the preciserecognition and positioning system of cotton heterogeneity fiber based on RBF-GMDH neuralnetwork. The system includes five modules: acquisition module, image pre-processing module,feature extraction module, identification module and positioning module. The main researchcontents and innovations of this paper are as follows:(1) Improved thresholding method is proposed. Based on Otsu and local threshold, this paperproposes an improved thresholding method. Through comparing each pixel of the original imageto its3×3neighborhood pixels and the Otsu threshold of the image, respectively, and the errorrange is set to5pixels by repeated experiments, determine whether the pixel is background orimpurities. It achieves an effective division of cotton impurities.(2) Window positioning method is proposed. The shape of cotton impurities is irregular and thereis no reliable ways to locate irregular objects currently. This paper proposes a positioningmethod based on windows. By dividing the whole image into many small windows with thesame size, the position of impurity is transformed to the two-dimensional coordinates of thesmall windows where cotton impurities locate. It achieves accurate positioning of cottonimpurities. This method is simple, easy to store, and has high positioning accuracy.(3) Improved GLCM extraction method is proposed. The GLCM extraction has a strongrelationship with the direction selected. Because the cotton impurity position in the image israndom, it’s difficult to select a direction to extract the GLCM. By taking the sum of GLCM at0°,90°,45°and135°directions, and then normalized it as the GLCM of the image. It achieves the extraction of cotton image texture features.(4) RBF neural network based on GMDH clustering is proposed. RBF neural network hasadvantages of a simple structure and fast learning and it can approximate any nonlinear function.However, there is no reliable theory to determine the RBF basis function centers currently.GMDH clustering can automatically determine the optimal number of clusters. This paperapplies GMDH clustering into RBF neural network to determine the optimal number of hiddennodes and then take the average of all types of samples as the corresponding basis functioncenters. It achieves rapid and accurate identification of cotton impurities.