Texture representation and analysis in material classification and characterization

Texture representation;Texture characterization;Deep learning;Texture classification;Texture tracking;Texture analysis;Texture dataset;Material recognition;Material surface characterization

Hu, Yuting ; AlRegib, Ghassan Electrical and Computer Engineering Anderson, David Jayaraman, Sundaresan Zhang, Ying Kalidindi, Surya ; AlRegib, Ghassan

University:Georgia Institute of Technology

Department:Electrical and Computer Engineering

关键词: Texture representation;    Texture characterization;    Deep learning;    Texture classification;    Texture tracking;    Texture analysis;    Texture dataset;    Material recognition;    Material surface characterization;   

Others  :  https://smartech.gatech.edu/bitstream/1853/61735/1/HU-DISSERTATION-2019.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

Objects and scenes in the real world exhibit abundant textural information. Textures observed in the surfaces of natural objects are not only the appearance of various materials, but also contain important visual cues (i.e., patterns and their spatial organizations). Recognition and characterization of textures and materials from their visual appearance has been effective in object recognition, robotics, quality inspection, scene understanding, facial image analysis, and medical or geological structure interpretation. In the industry, modern automated manufacturing systems utilize visual information to perform specific tasks such as classification, surface characterization, and visual tracking of products. However, these tasks are commonly conducted on products at the macroscopic scale with simple structures. When dealing with products at the microscopic scale (i.e., the fine-grained scale), describing and distinguishing similarmaterials with different texture characteristics such as smoothness poses a challenging problem, which requires discriminative and efficient texture representation techniques.In this dissertation, we develop texture representation algorithms in a progressive manner evolving from deterministic (i.e. handcrafted) to learning-based ones in order to handle both macroscopic and microscopic information. By involving high-order derivative features, multi-scale analysis, multi-level texture encoding, an end-to-end learning architecture, or geometric constraints in algorithm design, our texture representation methods enhance their discriminative representation ability and perform well on three main vision-based tasks, texture recognition, material surface characterization, and motion tracking of textures. Experimental results on typical texture and material datasets and our collected dataset show that our texture representation methods outperform their state-of-the-art counterparts in terms of accuracy and efficiency, thus aiding in the deployment of automated manufacturing systems.