Dr. Bugao Xu received his Ph. D. from the University of Maryland at College Park in August 1992, and started his career at the University of Texas at Austin as an Assistant Professor in January 1993. After 23 years of service at UT Austin, he joined the faculty of the University of North Texas (UNT) in 2016. Currently, he is a professor in the Department of Merchandising and Digital Retailing of UNT. His investigative research was focused on digital modeling and characterization of textile materials, 3D body imaging, virtual clothing, apparel fit and customization, made-to-measure technology, and AI in retail. He is also interested in cross-disciplinary research, such as obesity assessment and automated pavement inspection. He has authored over 220 refereed journal papers (>170 in SCI/SSCI journals) and more than 90 peer-reviewed conference proceedings. As a PI, he received a total of $5.3 million grants from NSF, USDA, NIH, Texas Higher Education Coordinating Board, Texas Food and Fiber Commission, Texas Department of Transportation, Cotton Inc., Walmart Foundation, VF and other private businesses. He also provided consultative or cooperative work on digital textiles to many companies such as P&G, Unilever, J&J, Maytag, and Cotton, Inc.
This project is focused on developing improved methods to assess phenotypic traits of cotton, with the ultimate goal of improving cotton fiber quality. Existing techniques to phenotype cotton fibers have a range of limitations, including inaccuracy, laborious sample preparation, and low throughput analysis. To address these challenges, our project aims to pioneer a new, high-throughput methodology for fiber phenotyping, encompassing both sample preparation and analysis. This will involve the development of robust models using deep learning algorithms to generate 3D surface representations of fibers for accurate fineness and maturity measurements. Additionally, we plan to utilize cotton diversity panels to calibrate and validate our results.
On-Loom Fabric Defect Inspection Using Contact Image Sensor and Machine Learning
This proposed project aims to develop an on-loom fabric inspection system that is reliable, practical, and cost-effective, utilizing contact image sensors (CIS) and advanced defect detection algorithms. Unlike traditional camera systems, which require separate lighting and stand at a distance, a CIS unit integrates the sensor chip, signal amplifier, and LED lights on the same substrate. This compact and lightweight design, along with an extremely short focal length, makes CIS particularly suitable for on-loom inspection where space, resolution, and cost are critical factors for success. Upon completion, the system will demonstrate early detection of defects, compact size without obstructing the weaver’s view, sub-millimeter resolution, and robust AI models for consistent defect detection.
Web mining system for retail analytics
We used big-data technology to create a scalable system that can fetch webpage contents of an online retailer, including customers’ reviews, ratings, demographic data and product information. The system can access and process not only HTML/CSS documents but also dynamic scripts and other resources in real-time. The downloaded contents can be used customer segmentation to enhance marketing strategy and customer loyalty programs. NLP techniques are used to construct deep-learning models to perform aspect-based sentiment analysis of user review texts for personalized product recommendations.
Fiber Image Analysis System--FIAS (US patent 7588438)
The FIAS is a customized microscopic image analysis system for automatic, high-volume measurements of fiber geometric attributes. The major components of the FIAS include a zoomable microscope, a high resolution camera, a motorized stage, a pneumatic fiber cutter/spreader, and image analysis software. FIAS offers two basic functions: longitudinal analysis and cross-sectional analysis of fibers.
The longitudinal analysis scans a fiber transversely along its axis, and provides the measurements of fiber diameter/width, diameter distribution, blend ratio of two different fibers, cotton maturity, cotton dead fiber ratio, etc. Prior to the analysis, fibers are cut into 0.5mm snippets, and spread onto a microscope slide using the pneumatic fiber cutter/spreader. The fiber snippets on the slide are consecutively imaged and measured as the slide is automatically shifted by the motorized stage. Proprietary computer algorithms include fiber axis calculation, double transverse scans, false scan prevention, dead fiber detection, and cotton maturity calculation. It is expected to measure up to 50,000 fiber scans per slide, which are sufficient for reliable statistic analysis.
The cross-sectional analysis provides feature measurements including fiber perimeter, area, equivalent diameter, and shape factors (circularity). For hollow fibers, such as cotton and medullated animal fibers, the wall thickness, maturity or medullation can be measured. For multi-lobal fibers, such as trilobal nylon fibers, the modification ratio can be measured.
FIAS (left) and Fiber Cutter/Spreader (right)
Imaging Colorimeter for Cotton Trash and Color Measurement
A new system that automatically captures images of raw cotton, locates non-lint particles in the image, and uses fuzzy logic and neural network methods to classify cotton colors and trash particles. The system considerably improves the agreements with the human classer. It can perform multiple functions: 1) trash content and grading; 2) trash category (leaf, bark and seedcoat); 3) content of yellow spot; 4) color data in various color spaces (Rdab, CIE L*a*b* or CIE L*c*h*) and color grading.
Objective Evaluation of Fabric Appearance
A “single hardware setup” solution was developed for multiple functions of objectively characterizing and evaluating surface changes of fabrics such as wrinkling, fuzzing, pilling and shrinking. The ratings of these surface features were trained using the standard replicas (AATCC and ASTM) or physical samples representing different scales set visually by experts. The system has greatly improved the objectivity of fabric appearance evaluations, and reduced the cost and the time for performing multiple tests.
Laser Profilometer for Surface Wear Assessment
The project was to measure surface characteristics (profile, roughness, fuzziness, etc.) of a wide range of materials such as fabric, plastic and metal. The system consists of a laser displacement sensor (4-um depth resolution), motorized x-y mechanical stage (70x70 mm2 travel), and customized software.
Three-Dimensional Body Imaging
For the past ten years, Dr. Bugao Xu and his “Human Dimension Research” team has been working on developing three dedicated 3D systems for fast and accurate generations of high-fidelity surface images.
(1) Rotary laser scanner
The rotary scanning unit mainly comprises a laser sensing unit and a step motor, which are installed on the top of a tripod. The computer then traces the line in each image and calculates the 3D coordinates of all the pixels on the line. The coordinates from the frontal and rear images are registered together to form a full body surface according to the relative positions of the two scanning units. A total of 250 to 350 images are captured and processed within 2-3 seconds, depending on the height of the subject. The rotary scanning makes the unit much more compact, lighter and thus more portable. Each unit weighs less than 2.5 kilograms, and the entire setup takes a footprint of 1.5m by 2.4m.
(2) Kinect body imaging system (KBI)
KBI system is a 3D Body Imaging and measurement system. The KBI uses 4 Microsoft Kinect sensors to acquire 3D data clouds of front and back surfaces of a whole body. After the precise registration, the front and back views of the body can be merged and reconstructed to form a smooth and complete digital model that is scalable and rotatable on the computer screen. The digital model can be used for measurement extractions, custom clothing and virtual try-on. The KBI system provides both hardware and software for 3D body imaging, which is quick (< 1s imaging time), portable (on stands), compact (2m x 1.5m footprint), economical and accurate (~1mm). It has a viewing volume of 2m x 1m x 0.8m (Height x Width x Depth). The body measurements include many circumference, length and volume data at pre-defined landmarks, and can be taken manually with “e-tape” tools.
(3) Stereovision body imaging system (SBI)
Stereovision Body Imaging (SBI) system uses 4 high-resolution digital cameras to take front and back images of a body, and a multi-resolution stereo-matching algorithm to generate 3D data clouds of the surface. After the precise registration, the front and back views of the body can be merged and reconstructed using the subdivision modeling algorithm to form a smooth and complete digital model that is scalable and rotatable on the computer screen. The SBI system solution for 3D body imaging is quick (0.2 s imaging time), portable (on stands), easy to calibrate, economical and accurate (<1mm). The scanned body images and their measurements can be stored in a database for later secure access. Possible applications include: obesity, body shape analysis and weight management, apparel fit and sizing system, and ergonomic products.
Virtual Clothing
The garment dressing simulation conducts a transformation of multiple 2D garment patterns into a 3D configuration that follows the surface of a human body, and displays the draping style of the garment. During this transformation, the dimensional stability of the garment patterns is an important constraint that needs to be imposed to ensure the created garment to have a size complied with the original design of the patterns. The four issues are fabric modeling, mesh generation, wrapping and draping.
Fabric modeling is a way to simulate fabric properties using a simple mechanic system. This is an important step for obtaining realistic visualization of a 3D garment. I A mass-spring system constitutes a matrix of mass particles, representing the distributed mass of the fabric, and three types of springs connecting the particles and representing three different internal forces: tensile, shearing and bending. A mass-spring system with reasonable combination of the stiffness and viscosity of the spring can develop a cloth-like deformational behavior.
Mesh generation is a process to automatically setup the distributions and connections of particles and springs on a given garment pattern, which may have irregular shapes. In our approach, this procedure is carried out in two steps: the first step is to generate the nodes for the given pattern that represent the mass distribution of the mass-spring system, and the second step is to generate a list of triangles in terms of Delaunay triangulation.
Wrapping is the 2D-3D transformation that merges the particles on the sewing edges of a pattern with those of the matching pattern, and moves the inner particles of all the sewn patterns towards the body surface. The stiffness coefficients of the tensile and shearing springs can be set purposely to a value much higher than their empirical values. In the wrapping, the gravity, friction, air resistance and other external forces are not considered, and the sewing forces are formed by an adaptive force field distributed on each particle. In each iteration of computing the displacements of the particles, the collisions between the particles and the body surface are detected to prevent surface penetrations. The wrapping ends when all the sewing edges are merged. Wrapping creates an initial status for draping.
Draping is the further pattern transformation to produce the draping effects by using realistic spring values in the fabric model, and adding all the external forces. Collision detection, and strain control and size stability are the major issues in the draping simulation.
Stereovision System for 3D imaging and Measurement
The prototype of a stereo vision system includes two CMOS digital cameras and a DPL digital projector. A random noise pattern is projected onto the object surface, and two digital images are acquired with one snapshot. We have implemented various effective stereo matching algorithms which can infer reliable disparities with sub-pixel accuracy, and the 3D modeling algorithms which can smoothen and simplify the raw data while keeping enough details.
High-Speed, Real-Time Highway Pavement Distress Inspection (US patent 7697727)
The system, equipped with a line-scan camera and a high-speed frame grabber, is installed in a designated vehicle, and can take up to 44,000 lines or 88 meters pavement surface images per second while it covers full pavement lane (3.66 meters). A multiresolution segmentation algorithm for crack detection was developed to meet the high-speed requirement. The algorithm takes less than 20 milliseconds as running on a 2 GHz Pentium IV processor and can reliably detect a variety of cracking distresses. The system is able to conduct the survey at a vehicle speed from 5 to 112 kmh-1, and report the cracking data in the ASSHTO, ASTM and PMIS formats and output a crackmap for the survey pavement. No human interference is required during the entire survey operation. The inspection data from the system have much better repeatability than those from the visual inspections.
A survey vehicle
A 60-feet pavement section (top) and the crackmap (bottom)
3D Pavement Scanning System
We developed an automated pavement inspection system using a 3D camera and a structured laser light to acquire dense transverse profiles of a pavement lane surface when it is carried by a moving vehicle. After the calibration, the 3D system can yield a depth resolution of 0.5 mm and a transverse resolution of 1.5 mm/pixel at the 1.4 m camera height from the ground. The scanning rate of the camera can be set to its maximum at 5000 lines/sec, allowing the density of scanned profiles to vary with the vehicle’s speed. The paper then illustrates the algorithms that utilize 3D information to detect pavement distresses, such as transverse, longitudinal and alligator cracking, and presents the field tests on the system’s repeatability when used to scan a sample pavement in multiple runs at the same vehicle speed, at different vehicle speeds and under different weather conditions. The results show that this dedicated 3D system can capture accurate pavement images that detail surface distress, and obtain consistent crack measurements in repeated tests and under different driving and lighting conditions.
3D imaging system to investigate associations between body shape and obesity
Currently, 33.6% adults are overweight and 34.9% are obese in the United States1. Excess body fat is a public health concern due to its association with diabetes mellitus type 2 and other diseases. It is not only the amount of fat, but also its distribution in the body that is related to disease risks2–4. A meta-analysis by the authors found that diabetic/pre-diabetic populations have a great degree of visceral (abdominal) adiposity as compared to non-diabetic populations5. Thus, visceral adiposity (central obesity) as opposed to subcutaneous depots has greater implications in terms of diabetic risk. A new form of 3D imaging (stereovision) has enabled us to create more accurate, consistent and comprehensive body measures6. We have developed image-processing algorithms for unsupervised assessment of visceral and subcutaneous adipose tissue volumes from MRI abdominal images (Fig. 1), designed novel body shape descriptors to characterize overall and central obesity using 3D images from stereovision, and investigated the associations between internal adiposity data (MRI) and external body shape measurements (3D images). The new approach permits a more cost-effective and accurate obesity assessment for public health research.
Papers Published in SCI/SSCI Journals (*Corresponding author; +Supervised student, visiting scholar or postdoc researcher.)