MULTI-ILLUMINATION OPTICAL DESCRIPTORS IN DUAL-BRANCH SVM APPROACH IN CLASSIFYING EDGES OF MATERIALS DEPLOYED ON EDGES

Abstract

The issue of automated fabric recognition in the textile manufacturing industry is still a challenge because it is difficult to differentiate between substances of the same appearance yet having different physical characteristics. Whereas the weave patterns can be determined with the help of the geometric analysis, the recognition of the material composition demands more advanced techniques. The current paper introduces a biomimetic visual analysis system based on the processes of human experts inspection. We suggest a two-branch model in which the analysis of the geometric structure is mixed with multi-oriented- illumination- optical feature extraction. With the help of controlled lighting conditions, i.e. direct, angled (45deg), and backlight illumination, we pull out a set of specular reflection maps, surface roughness indices, as well as light absorption coefficients which are effective in distinguishing between natural and synthetic fibers. Our approach manages to identify materials with accuracy of 87.3% in cotton, polyester, silk and blended fabrics when geometry is used compared to 64.2% in geometry only. The system is based on edge computing hardware (ESP32-CAM), which allows connecting in real-time with industrial sewing machines. It is experimentally validated by using 1,200 samples of fabric, which shows good performance in different colors and densities of weaves.