A seminar I presented on using hyperspectral imaging to non-invasively phenotype prawns. The project focused on removing the background of the images, developing an intuitive UI and applying statistical algorithms. The aim was to train a model to predict the biochemical attributes of prawns using their spectra.
Hubble Asteroid Hunter III. Physical properties of newly found asteroids
Rapid phenotyping of prawns using hyperspectral imaging
1. Rapid phenotyping of
prawn biochemical
attributes using
hyperspectral imaging
Stuart Hinchliff supervised by Professor Ronald White and Professor
Dean Jerry
2. Index
• Aims
• Background information
• Techniques applied and results
• Preliminary statistics
• UI demonstration
• Outcomes and future work
3. Motivation
• The investigation of hyperspectral imaging as a
fast and non-invasive technique that could lead
to improved selective prawn breeding programs.
4. Aims:
• To obtain hyperspectral image data on a variety of
prawns.
• To investigate appropriate image processing methods for
distinguishing prawns.
• To prepare a statistical environment for correlating
prawn spectra to biochemical attributes, and training
models.
• To structure the explored techniques in an intuitive user
interface.
5. Hyperspectral Imaging
• Typical RGB image (eg. “normal” image in formats such
as .jpeg) is made up of 3 bands:
• Multispectral images and hyperspectral images have many more than this
(our images have 240 bands ranging from ~ 400 nm to 1000 nm).
• Average image size: 800 MB
7. NIRS (Near-Infrared Spectroscopy)
• Higher energy than mid-IR and therefore is useful in
probing bulk material with little to no preparation (water
is reasonably transparent in NIR)
• Region of electromagnetic spectrum: 700nm to 2500nm
• Complex spectra due to molecular overtone (harmonics)
and combination vibrations – to extract chemical
information, multivariate calibration techniques are used
such as:
Principal component analysis,
Partial least squares, and
Neural networks.
8. Obtaining the spectra
• Remove undesirable background elements (tray, label and rubber
band)
• Two approaches: using traditional RGB methods, using
hyperspectral techniques
Conversion
Band
Label
9. RGB Techniques
• Considered due to low volume of data, therefore high speed and efficiency
• Thresholding and morphological operations:
• Advantages: Very efficient with reasonable accuracy
• Disadvantages: Different lighting conditions would require calibration of
threshholding, missing information.
Example 1 Example 2
10. RGB Techniques continued…
• Clustering:
• Advantages: Low use of resources, more robust than morphological
operations
• Disadvantages: Lower accuracy (doesn’t distinguish labels), more clusters
significantly increase computational time, clustering is randomised
Conversion to L*a*b space
before using kmeans
Conversion to chromaticity
space before using kmeans
11. Scyllarus
• Scyllarus is hyperspectral software developed by NICTA
• A C++ API and a MATLAB toolbox are available.
• Advantages: Uses advanced algorithms to pre-process images and
identify materials. These materials could be useful for identifying
trends.
• Disadvantages: Poor efficiency (perhaps C++ API could be used).
12. Neural Networks
• Supervised machine learning technique that adjusts weights to ensure
inputs match output
• Uses nested cross-validation (training, test and validation data) to optimise
algorithm for the data and avoid overfitting
13. Neural Networks
• Supervised training on a manually classified image.
• Advantages: Very accurate, is robust and can be improved with further training
• Disadvantages: Not as efficient as other methods, noisy pixels will be
misclassified
14. Statistics and Analysis
• Environment developed to obtain prawn spectra signatures from images,
visualise trends (principal component analysis), preprocessing (scaling) and
train on a factor (neural networks).
15. Statistics and Analysis
• The investigation has demonstrated that no trends exist
for training on area, however we are optimistic for other
parameters.
17. Outcomes
• Software has been developed as an all-in-one
package for reading data, identifying prawns and
analysing trends.
• Scyllarus uses advanced preprocessing to identify
different materials in the images – could be
extremely useful in analysing segments.
• Solid foundation for future work. Could be a
suitable honours project.
18. Future Work
• Training on the biochemical data.
• Additional settings and features for user to
tweak to aid in training/visualising trends.
• Training on a particular wavelength could
remove the need of slow hyperspectral imaging
process.
• Perhaps switching languages and using the C++
API due to its improved efficiency.