A Distributed System-Based Multiplex Networks to Extract Texture Feature

Introduction

Appearance defect detection is an important part of ex-factory quality detection in manufacturing. During the production process, the factory must strictly detect the castings to ensure appearance quality. The surface condition of the casting depends on the equipment, pouring, and molding. With the advantages of high speed and accuracy, machine vision gradually replacing the low-efficiency manual sampling detection and has been widely used in automatic detection. The variety of surface roughness and defect types will increase the difficulty of adaptive defect detection. The standard of artificial discrimination of casting defects mainly depends on the texture. The importance of texture in human vision is one of the essential steps towards classifying objects and recognizing a scene. For example, humans can directly perceive defective regions where roughness changes from the neighborhood. The appearance defects of castings are caused by different reasons, resulting in different patterns of manifestation. The surface of the casting does not have the quasi-periodic characteristics of the machined surfaces. The surface of the casting is rough, showing the form of a mountain range after infinite magnification. That is, the texture is the integration of various mountains.

Texture analysis is a classic tool in defect recognition. The texture is a visual pattern that repeats in space (Liu Li et al. 2019). In other words, the texture is the spatial organization of a set of primitives. Although texture lacks a formal definition (Humeau-Heurtier Anne 2019), it does not affect the addition and research of texture analysis methods. The classic methods of texture analysis mainly include statistical approaches, structural approaches, spectral approaches. The Gray Level Cooccurrence matrix (Gotlieb Calvin C. & Kreyszig Herbert E 1990), Laws texture (Chung-Ming Wu et al. 1992), and local binary mode (LBPs) is a common tool in statistical methods. The basic idea of structural approaches is to describe texture in the form of appropriate statistical pattern recognition. The coordinate system of the image represented by the spectral approaches is closely related to the texture feature. The representative methods of spectral approaches include Fourier transform-based, wavelet-based, and Gabor-based method (Li Chaorong & Huang Yuanyuan 2017). Currently, there is no uniform manual feature descriptor for all types of texture. Besides, scholars have proposed model-based approaches and learning-based approaches in recent years, such as complex networks (Backes André Ricardo et al. 2013), gravitational models (De Mesquita Sá Junior Jarbas Joaci, et al. 2013), vocabulary learning, and deep learning methods (Ferreira Carlos A et al. 2019).

Human vision cannot discriminative between two textures with the same second-order statistics. Since the texture is a spatial phenomenon, the propose method not only consider isolated primitives for pattern analysis of local parts but also aggregate them into a global representation (Xu Degang et al. 2015). In this case, complex network theory is a powerful method that can reflect both local and global information of the image (Couto Leandro N et al. 2015). Complex networks can represent virtually any nature under their flexibility and versatility. Various studies represent a complex system as a complex network and analyze the topology and feature. The texture recognition method of a complex network is to obtain the centrality parameters of the nodes under the topology. The most popular method is to designate pixels as nodes of a complex network and use node similarity as a reference for edge weight based on Euclidean distance and node intensity (Scabini Leonardo F. S. et al. 2019 and Scabini Leonardo F. S. 2019).