Project 1. High-Throughput Phenotyping of Cotton Plants
Sponsor: Bayer CropScience
Phenotype measurements of a cotton plant (such as height and of number of cotton bolls) are currently carried out by hand. This project's goal is to allow for automatic detection and measurement of those features. To accomplish its goals, the project implements image processing and machine learning techniques. The images are captured by cameras mounted on a robot:
Below, the robot in action:
And some results of the cotton boll detection (second and third rows):
Project 2: Automatic Quail Population Density Estimation
Partner: Department of Natural Resources Management, Texas Tech University
This project's goal is to detect individual quails from audio recordings done in a
rural area of Lubbock, TX, estimating the local quail population. The detection of
quail calls can be seen on the plot below. The superimposed grey plot is the spectrogram
of the audio recording:
Quail calls detected in recording marked by the diamonds on amplitude peaks
The user interface of the software developed at AVL for detecting the quail calls:
User interface of the software developed to detect quail calls
Past Projects - below, a sample of past projects carried out at AVL
An IR-based Machine Vision System for Objective Quantification of Moisture Transport in Fabrics
Sponsor: Cotton Incorporated
System at work
The wicking characteristic of a fabric, i.e., the movement of water or liquid through a fabric, is an important functional attribute of a fabric, which necessitates the development of techniques to quantify moisture transport. This IR-based computer vision system uses image processing techniques to estimate wicking, static, and drying periods of moisture in cotton fabrics. The system captures a thermal image (infrared) of a piece of fabric. The algorithm finds a region of interest on the captured image and segments the area that contains moisture. Wicking, static, and drying periods are recorded, and Area-Time Profiles plots are created using those numbers. An example can be seen below:
After several runs of the experiment, a model for moisture transport behavior was established for control and treated fabric. The features created from the data were wicking rate, wicking duration, static average area, static duration, drying rate and drying duration:
a - c: Wicking; d: Static; e-f: Drying Periods
Cotton Fiber Quality Assessment Using High-Resolution Imaging
Sponsor: Cotton Incorporated
3D Reconstruction of cotton fiber using segmented slices
The objective of this project was to study and quantify length and maturity measurements for individual cotton fibers to obtain the bivariate distribution of sample length and maturity. Longitudinal scans of cotton fibers were acquired by a confocal microscopy system and processed using image segmentation algorithms. The segmented images were processed with a length algorithm and a maturity algorithm.
The length algorithm computed fiber length using the Digital Straight Segments algorithm. The maturity algorithm extracted several features from the segmented images, based on physical properties of the fiber and on the image's texture. Those features defined the target domain (the longitudinal scans mentioned before, where maturity values were not known), while previously recorded data about perimeter and area (fiber cross-sections, with known maturity values) of the cotton fibers was used as the source domain. To reduce differences in feature distributions between the source domain and the target domain, transfer learning techniques were implemented. This process of domain adaptation allowed us to correlate events between the two domains.
Then, machine learning algorithms were applied to relate the two domains and to quantify fiber maturity o the target domain. It was observed that, within a single cotton fiber, maturity can vary significantly (>50%). This was confirmed with visual inspection as well as quantitative measurements.
Analysis and visualization of reconstructed immature cotton fibers
Analysis and visualization of reconstructed mature cotton fibers
IT-SNAPS: A GPU-based, User-Aware Interactive Texture Segmentation System
We have designed and developed a new framework for an interactive texture segmentation technique, which we term as IT-SNAPS (Interactive Texture-Snapping System). Unlike its predecessors, IT-SNAPS can effectively aid the user in accurately segmenting images with complex texture, without placing undue burden on the user. This is made possible through the formulation of IT-SNAPS, which enables it to be user-aware, i.e., it unobtrusively elicits information from the user during the segmentation process, and hence, adapts itself on-the-fly to the boundary being segmented. In addition to generating an accurate segmentation, IT-SNAPS allows for extraction of useful information post-segmentation, which can potentially assist in the development of customized automatic segmentation algorithms. The core computational engine of IT-SNAPS has been designed to operate on NVIDIA's GPU, which enables IT-SNAPS to be rapidly-trainable, thus, greatly increasing its efficiency as a interactive delineation tool.
GPU-powered Machine Vision System for Simultaneous Quantification of Shrinkage and Stain Release on Solid and Patterned Fabrics
We have developed a machine vision system that can automatically detect and recognize multiple objects in the presence of a heavily textured and noisy background. The objects of interest in this application are shrinkage dots and stains, which, once detected and localized, can yield important functional measurements of a fabric. Termed “shrinkage” and “stain release”, these functional attributes impact the fabric pricing, and hence, objective measurements of these parameters are of significant consequence to the textile industry.
The proposed vision-based approach operates within a semi-supervised framework, wherein it is assumed that the machine vision system has the knowledge of the background texture/pattern. Using a customized combination of area-based and feature-based registration techniques, the objects of interest are enhanced with respect to the background, following which, the stains are localized using a GPU-based implementation of adaptive statistical snakes.
Integrating Internal Properties into an Interactive Segmentation Framework
Incorporating shape-based information into an ‘on-the-fly' interactive segmentation scheme is a challenging task due to the unavailability of the complete shape information during the segmentation procedure. A partial-shape registration technique in combination with shape-based level sets is employed in order to integrate internal properties, such as shape and symmetry into an interactive segmentation framework.
Morphometry of Cervical and Lumbar Vertebrae
We continue to develop an extensive tool under Matlab that is capable of segmenting and characterizing the lumbar and cervical vertebrae from digitized x-ray images of the spine. The tool is comprised of customized shape-based segmentation algorithms.
Detecting Biliary Dyskinesia from 3D Ultrasound Images
Current medical analysis of proper gallbladder function relies on the measurement of volume change over time. This project seeks to replace the current method of measurement, which requires the injection of radioactive markers for imaging, with a non-invasive 3D ultrasound approach. A 3D segmentation toolbox has been built in order to segment the gallbladder and subsequently detect biliary dyskinesia.
On Loom Defect Detection
The On-Loom Defect Detection houses a robust defect detection algorithm while overcoming vibration, variable loom speed, and nonuniform sensor response. I was the first on-loom fabric inspection system created, using a DSP-based board to acquire high quality images ranging from 100-200 pixels/inch.
The algorithm takes the image through the processes displayed in the flowchart below:
On Loom Defect Detection algorithm flowchart
The response of Multiscale Wavelet Representation is optimized for appropriate basis functions and number of iterations. The output images are then fused and assessed for homogeneity, as it was observed that defects reduce the global homogeneity. After the image undergoes these processes, optimal threshold and blob analysis are used to identify or localize the defect, as shown below:
Example of the algorithm at work
The learning algorithm was optimized to detect 17 different filament yarn defects and 26 spun yarn defects. More than 3,700 fabric images were acquired and processed. The defect detection rate (true positives) reached roughly 89%, while false positives were detected on 2.5% of the images.
Part-Based Object Recognition for Aerial Image Synthesis and Analysis
Complicated objects can generally be decomposed into smaller building blocks (parts). Understanding contextual relations between these parts in modelling an object recognition system is of great importance for various applications. This project involves using the state-of-the-art, part-based algorithms, such as Conditional Random Fields and Parse Graphs, to aid in image synthesis and object recognition.
Plant Water Stress Estimation through Processing of IR and Color Images
We are developing an automated vision system that estimates plant water stress status by utilizing infrared and color images. The system will be applied and validated in a study of cotton plant water stress tolerance in the AgriLife Center in Lubbock.