关键词: 高分子滤网, 机器视觉, 疵点检测, 局部二值模式, 支持向量机

Abstract: The spiral mesh is a kind of polymer filter, which is formed by polyester, polyamide, polyethylene and other polymer monofilaments through the winding device to form a left and right opposite spiral ring line similar to the spring structure. Through the left and right spiral ring lines, the mesh is formed by repeatedly meshing and inserting the connecting core wire in the overlapping part, and then the spiral mesh is finally formed by the subsequent processes such as heat setting, core insertion, cutting, edge sealing and gluing. It has special void structure, high interception precision, high mesh flatness, free splicing, high structural strength, long service life and strong corrosion resistance, and can realize solid-liquid separation in various scenarios. It is widely used in environmental protection, papermaking, coal mine, food, medicine and other fields. Spiral mesh defects are mainly divided into three categories: holes, missing core wires and staggered rings. Although most of its production processes have been automated, the final quality inspection process still depends on manual inspection. However, manual detection has shortcomings such as strong subjectivity, high cost, low efficiency, high labor intensity, and damage to vision, which cannot guarantee the detection precision and accuracy. The speed and accuracy of manual detection can no longer meet the requirements of spiral mesh production. It is particularly important to use machine vision to realize the automation of spiral mesh defect detection.
In order to solve the problems of low efficiency and high false detection rate of spiral network manual defect detection, we proposed a method of spiral network defect detection based on classification. By extracting multi-mode and multi-scale LBP features, the information of spiral network image was fully characterized. Based on the classification idea, in the case of small sample size, the SVM classifier was used to effectively learn and classify the small sample, high-dimensional and non-linear data, and to distinguish the local damaged and abnormal defective images and defect-free images to realize the automatic detection of spiral mesh defects. By building an image acquisition device, we collected the defective and non-defective images of the spiral mesh, and verified the proposed method. By extracting the texture features of the spiral mesh, we discussed the classification effects of LBP features of different modes and scales and other features on SVM and K-NN classifiers. It is found that the most suitable feature for spiral mesh defect detection is the uniform mode LBP operator with eight sampling points and two sampling radius, and the optimal classifier is SVM. The effectiveness of the proposed method was verified by collecting images to establish a data set. The experimental results show that the classification accuracy of the spiral network with and without defects reaches 100 %, which verifies that the LBP operator has good anti-noise and robustness. The average classification speed is 0.48 s/sheet.
This paper proposes a comprehensive and efficient spiral network defect detection algorithm, which has certain practical significance for the spiral network industry. The sample size of the spiral network defect images collected in this paper is small. On this basis, the number of samples will be further increased to establish a more stable spiral network defect detection system.

Key words: polymer filter, machine vision, defect detection, local binary pattern, support vector machine